voila

TunisianASR/results/14epoch_tunisian/1234/app.py

@@ -0,0 +1,772 @@
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+from cv_train import ASRCV
+import torchaudio
+import numpy as np
+import kenlm
+from pyctcdecode import build_ctcdecoder
+import re
+from torch.nn.utils.rnn import pad_sequence
+import torch.optim as optim
+import torch.nn as nn
+
+
+# Commented out IPython magic to ensure Python compatibility.
+hparams_file, run_opts, overrides = sb.parse_arguments(["TunisianASR/semi_trained.yaml"])
+
+# If distributed_launch=True then
+# create ddp_group with the right communication protocol
+sb.utils.distributed.ddp_init_group(run_opts)
+
+with open(hparams_file) as fin:
+    hparams = load_hyperpyyaml(fin, overrides)
+
+# Create experiment directory
+sb.create_experiment_directory(
+    experiment_directory=hparams["output_folder"],
+    hyperparams_to_save=hparams_file,
+    overrides=overrides,
+)
+# Dataset prep (parsing Librispeech)
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+
+    # 1. Define datasets
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(
+            sort_key="duration",
+            key_max_value={"duration": hparams["avoid_if_longer_than"]},
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["dataloader_options"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration",
+            reverse=True,
+            key_max_value={"duration": hparams["avoid_if_longer_than"]},
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["dataloader_options"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    # We also sort the validation data so it is faster to validate
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav):
+        info = torchaudio.info(wav)
+        sig = sb.dataio.dataio.read_audio(wav)
+        if len(sig.shape)>1 :
+            sig = torch.mean(sig, dim=1)
+        resampled = torchaudio.transforms.Resample(
+            info.sample_rate, hparams["sample_rate"],
+        )(sig)
+        return resampled
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "blank_label": hparams["blank_index"],
+        "unk_label": hparams["unk_index"]
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets, ["id", "sig", "wrd", "char_list", "tokens"],
+    )
+    return train_data, valid_data,test_datasets, label_encoder
+
+class ASR(sb.core.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        if stage == sb.Stage.TRAIN:
+            if hasattr(self.hparams, "augmentation"):
+                wavs = self.hparams.augmentation(wavs, wav_lens)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs, wav_lens)
+        x = self.modules.enc(feats)
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+
+        return p_ctc, wav_lens
+
+    def custom_encode(self,wavs,wav_lens) :
+        wavs = wavs.to("cpu")
+        if(wav_lens is not None): wav_lens.to(self.device)
+
+        feats = self.modules.wav2vec2(wavs, wav_lens)
+        x = self.modules.enc(feats)
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+
+        return feats,p_ctc
+
+
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC) given predictions and targets."""
+
+        p_ctc, wav_lens = predictions
+
+        ids = batch.id
+        tokens, tokens_lens = batch.tokens
+
+        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+
+        if stage != sb.Stage.TRAIN:
+            predicted_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+            # Decode token terms to words
+            if self.hparams.use_language_modelling:
+                predicted_words = []
+                for logs in p_ctc:
+                    text = decoder.decode(logs.detach().cpu().numpy())
+                    predicted_words.append(text.split(" "))
+            else:
+                predicted_words = [
+                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                    for utt_seq in predicted_tokens
+                ]
+            # Convert indices to words
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        should_step = self.step % self.grad_accumulation_factor == 0
+        # Managing automatic mixed precision
+        # TOFIX: CTC fine-tuning currently is unstable
+        # This is certainly due to CTC being done in fp16 instead of fp32
+        if self.auto_mix_prec:
+            with torch.cuda.amp.autocast():
+                with self.no_sync():
+                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+            with self.no_sync(not should_step):
+                self.scaler.scale(
+                    loss / self.grad_accumulation_factor
+                ).backward()
+            if should_step:
+
+                if not self.hparams.wav2vec2.freeze:
+                    self.scaler.unscale_(self.wav2vec_optimizer)
+                self.scaler.unscale_(self.model_optimizer)
+                if self.check_gradients(loss):
+                    if not self.hparams.wav2vec2.freeze:
+                        if self.optimizer_step >= self.hparams.warmup_steps:
+                            self.scaler.step(self.wav2vec_optimizer)
+                    self.scaler.step(self.model_optimizer)
+                self.scaler.update()
+                self.zero_grad()
+                self.optimizer_step += 1
+        else:
+            # This is mandatory because HF models have a weird behavior with DDP
+            # on the forward pass
+            with self.no_sync():
+                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+
+            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+
+            with self.no_sync(not should_step):
+                (loss / self.grad_accumulation_factor).backward()
+            if should_step:
+                if self.check_gradients(loss):
+                    if not self.hparams.wav2vec2.freeze:
+                        if self.optimizer_step >= self.hparams.warmup_steps:
+                            self.wav2vec_optimizer.step()
+                    self.model_optimizer.step()
+                self.zero_grad()
+                self.optimizer_step += 1
+
+        self.on_fit_batch_end(batch, outputs, loss, should_step)
+        return loss.detach().cpu()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            if not self.hparams.wav2vec2.freeze:
+                sb.nnet.schedulers.update_learning_rate(
+                    self.wav2vec_optimizer, new_lr_wav2vec
+                )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+
+        # If the wav2vec encoder is unfrozen, we create the optimizer
+        if not self.hparams.wav2vec2.freeze:
+            self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+                self.modules.wav2vec2.parameters()
+            )
+            if self.checkpointer is not None:
+                self.checkpointer.add_recoverable(
+                    "wav2vec_opt", self.wav2vec_optimizer
+                )
+
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+    def zero_grad(self, set_to_none=False):
+        if not self.hparams.wav2vec2.freeze:
+            self.wav2vec_optimizer.zero_grad(set_to_none)
+        self.model_optimizer.zero_grad(set_to_none)
+
+
+from speechbrain.pretrained import EncoderASR,EncoderDecoderASR
+french_asr_model = EncoderASR.from_hparams(source="asr-wav2vec2-commonvoice-fr", savedir="pretrained_models/asr-wav2vec2-commonvoice-fr")
+french_asr_model.to("cpu")
+cvhparams_file, cvrun_opts, cvoverrides = sb.parse_arguments(["EnglishCV/train_en_with_wav2vec.yaml"])
+with open(cvhparams_file) as cvfin:
+    cvhparams = load_hyperpyyaml(cvfin, cvoverrides)
+cvrun_opts["device"]="cpu"
+english_asr_model = ASRCV(
+        modules=cvhparams["modules"],
+        hparams=cvhparams,
+        run_opts=cvrun_opts,
+        checkpointer=cvhparams["checkpointer"],
+    )
+english_asr_model.modules.to("cpu")
+english_asr_model.device="cpu"
+english_asr_model.checkpointer.recover_if_possible()
+run_opts["device"]="cpu"
+print("moving to tunisian model")
+asr_brain = ASR(
+    modules=hparams["modules"],
+    hparams=hparams,
+    run_opts=run_opts,
+    checkpointer=hparams["checkpointer"],
+)
+asr_brain.modules.to("cpu")
+asr_brain.checkpointer.recover_if_possible()
+asr_brain.modules.eval()
+english_asr_model.modules.eval()
+french_asr_model.mods.eval()
+asr_brain.modules.to("cpu")
+
+# Commented out IPython magic to ensure Python compatibility.
+# %ls
+
+#UTILS FUNCTIOJNS
+def get_size_dimensions(arr):
+    size_dimensions = []
+    while isinstance(arr, list):
+        size_dimensions.append(len(arr))
+        arr = arr[0]
+    return size_dimensions
+
+def scale_array(batch,n):
+    scaled_batch = []
+
+    for array in batch:
+        if(n < len(array)): raise ValueError("Cannot scale Array down")
+
+        repeat = round(n/len(array))+1
+        scaled_length_array= []
+
+        for i in array:
+            for j in range(repeat) :
+                if(len(scaled_length_array) == n): break
+                scaled_length_array.append(i)
+
+        scaled_batch.append(scaled_length_array)
+
+    return torch.tensor(scaled_batch)
+
+
+def load_paths(wavs_path):
+    waveforms = []
+    for path in wavs_path :
+        waveform, _ = torchaudio.load(path)
+        waveforms.append(waveform.squeeze(0))
+    # normalize array length to the bigger arrays by pading with 0's
+    padded_arrays = pad_sequence(waveforms, batch_first=True)
+    return torch.tensor(padded_arrays)
+
+
+
+device = 'cpu'
+verbose = 0
+#FLOW LEVEL FUNCTIONS
+def merge_strategy(embeddings1, embeddings2, embeddings3,post1, post2,post3):
+
+
+    post1 = post1.to(device)
+    post2 = post2.to(device)
+    post3 = post3.to(device)
+    embeddings1 = embeddings1.to(device)
+    embeddings2 = embeddings2.to(device)
+    embeddings3 = embeddings3.to(device)
+
+    posteriograms_merged = torch.cat((post1,post2,post3),dim=2)
+    embeddings_merged = torch.cat((embeddings1,embeddings2,embeddings3),dim=2)
+
+    if(verbose !=0):
+      print('MERGED POST ',posteriograms_merged.shape)
+      print('MERGED emb ',embeddings_merged.shape)
+
+    return torch.cat((posteriograms_merged,embeddings_merged),dim=2).to(device)
+
+def decode(model,wavs,wav_lens):
+
+    with torch.no_grad():
+        wav_lens = wav_lens.to(model.device)
+        encoder_out = model.encode_batch(wavs, wav_lens)
+        predictions = model.decoding_function(encoder_out, wav_lens)
+        return predictions
+
+def middle_layer(batch, lens):
+
+    tn_embeddings, tn_posteriogram = asr_brain.custom_encode(batch,None)
+
+    fr_embeddings = french_asr_model.mods.encoder.wav2vec2(batch)
+    fr_posteriogram =french_asr_model.encode_batch(batch,lens)
+    en_embeddings = english_asr_model.modules.wav2vec2(batch, lens)
+    x = english_asr_model.modules.enc(en_embeddings)
+    en_posteriogram = english_asr_model.modules.ctc_lin(x)
+    #scores, en_posteriogram = english_asr_model.mods.decoder(en_embeddings ,lens)
+    if(verbose !=0):
+      print('[EMBEDDINGS] FR:',fr_embeddings.shape, "EN:",en_embeddings.shape, "TN:", tn_embeddings.shape)
+      print('[POSTERIOGRAM] FR:',fr_posteriogram.shape, "EN:",en_posteriogram.shape,"TN:",tn_posteriogram.shape)
+
+
+    bilangual_sample = merge_strategy(fr_embeddings,en_embeddings,tn_embeddings,fr_posteriogram,en_posteriogram,tn_posteriogram)
+    return bilangual_sample
+
+class Mixer(sb.core.Brain):
+
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        wavs, wav_lens = batch.sig
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        if stage == sb.Stage.TRAIN:
+            if hasattr(self.hparams, "augmentation"):
+                wavs = self.hparams.augmentation(wavs, wav_lens)
+
+        multi_langual_feats = middle_layer(wavs, wav_lens)
+        multi_langual_feats= multi_langual_feats.to(device)
+        feats, _ = self.modules.enc(multi_langual_feats)
+        logits = self.modules.ctc_lin(feats)
+        p_ctc = self.hparams.log_softmax(logits)
+        
+        if stage!= sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        else : 
+            p_tokens = None
+        return p_ctc, wav_lens, p_tokens
+    
+    
+    def treat_wav(self,sig):
+        multi_langual_feats = middle_layer(sig.to("cpu"), torch.tensor([1]).to("cpu"))
+        multi_langual_feats= multi_langual_feats.to(device)
+        feats, _ = self.modules.enc(multi_langual_feats)
+        logits = self.modules.ctc_lin(feats)
+        p_ctc = self.hparams.log_softmax(logits)
+        predicted_words =[]
+        for logs in p_ctc:
+            text = decoder.decode(logs.detach().cpu().numpy())
+            predicted_words.append(text.split(" "))
+        return " ".join(predicted_words[0])
+        
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC) given predictions and targets."""
+
+        p_ctc, wav_lens , predicted_tokens= predictions
+
+        ids = batch.id
+        tokens, tokens_lens = batch.tokens
+
+        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+
+
+        if stage == sb.Stage.VALID:
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+        if stage ==sb.Stage.TEST : 
+            if self.hparams.language_modelling:
+                predicted_words = []
+                for logs in p_ctc:
+                    text = decoder.decode(logs.detach().cpu().numpy())
+                    predicted_words.append(text.split(" "))
+            else : 
+                predicted_words = [
+                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                    for utt_seq in predicted_tokens
+                ]
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        should_step = self.step % self.grad_accumulation_factor == 0
+        # Managing automatic mixed precision
+        # TOFIX: CTC fine-tuning currently is unstable
+        # This is certainly due to CTC being done in fp16 instead of fp32
+        if self.auto_mix_prec:
+            with torch.cuda.amp.autocast():
+                with self.no_sync():
+                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+            with self.no_sync(not should_step):
+                self.scaler.scale(
+                    loss / self.grad_accumulation_factor
+                ).backward()
+            if should_step:
+
+
+                self.scaler.unscale_(self.model_optimizer)
+                if self.check_gradients(loss):
+                    self.scaler.step(self.model_optimizer)
+                self.scaler.update()
+                self.zero_grad()
+                self.optimizer_step += 1
+        else:
+            # This is mandatory because HF models have a weird behavior with DDP
+            # on the forward pass
+            with self.no_sync():
+                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+
+            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+
+            with self.no_sync(not should_step):
+                (loss / self.grad_accumulation_factor).backward()
+            if should_step:
+                if self.check_gradients(loss):
+                    self.model_optimizer.step()
+                self.zero_grad()
+                self.optimizer_step += 1
+
+        self.on_fit_batch_end(batch, outputs, loss, should_step)
+        return loss.detach().cpu()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+    def zero_grad(self, set_to_none=False):
+
+        self.model_optimizer.zero_grad(set_to_none)
+
+
+
+
+hparams_file, run_opts, overrides = sb.parse_arguments(["cs.yaml"])
+
+# If distributed_launch=True then
+# create ddp_group with the right communication protocol
+sb.utils.distributed.ddp_init_group(run_opts)
+
+with open(hparams_file) as fin:
+    hparams = load_hyperpyyaml(fin, overrides)
+
+# Create experiment directory
+sb.create_experiment_directory(
+    experiment_directory=hparams["output_folder"],
+    hyperparams_to_save=hparams_file,
+    overrides=overrides,
+)
+def read_labels_file(labels_file):
+    with open(labels_file, "r",encoding="utf-8") as lf:
+        lines = lf.read().splitlines()
+        division = "==="
+        numbers = {}
+        for line in lines :
+            if division in line :
+                break
+            string, number = line.split("=>")
+            number = int(number)
+            string = string[1:-2]
+            numbers[number] = string
+        return [numbers[x] for x in range(len(numbers))]
+
+label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+special_labels = {
+    "blank_label": hparams["blank_index"],
+    "unk_label": hparams["unk_index"]
+}
+label_encoder.load_or_create(
+    path=lab_enc_file,
+    from_didatasets=[[]],
+    output_key="char_list",
+    special_labels=special_labels,
+    sequence_input=True,
+)
+
+
+labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+labels = [""] + labels[1:-1] + ["1"]
+if hparams["language_modelling"]:
+    decoder = build_ctcdecoder(
+        labels,
+        kenlm_model_path=hparams["ngram_lm_path"],  # either .arpa or .bin file
+        alpha=0.5,  # tuned on a val set
+        beta=1,  # tuned on a val set
+    )
+
+
+
+run_opts["device"]="cpu"
+
+mixer = Mixer(
+    modules=hparams["modules"],
+    hparams=hparams,
+    run_opts=run_opts,
+    checkpointer=hparams["checkpointer"],
+)
+mixer.tokenizer = label_encoder
+mixer.device = "cpu"
+mixer.checkpointer.recover_if_possible()
+mixer.modules.eval()
+
+
+label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+
+# We dynamicaly add the tokenizer to our brain class.
+# NB: This tokenizer corresponds to the one used for the LM!!
+
+decoder = build_ctcdecoder(
+    labels,
+    kenlm_model_path= "arpas/everything.arpa",  # either .arpa or .bin file
+    alpha=0.5,  # tuned on a val set
+    beta=1,  # tuned on a val set
+)
+
+
+
+device = "cpu"
+mixer.device= "cpu"
+mixer.modules.to("cpu")
+
+from enum import Enum, auto
+class Stage(Enum):
+    TRAIN = auto()
+    VALID = auto()
+    TEST = auto()
+
+asr_brain.on_evaluate_start()
+asr_brain.modules.eval()
+
+
+import gradio as gr
+
+def treat_wav_file(file_mic,file_upload ,asr=mixer, device="cpu") :
+    if (file_mic is not None) and (file_upload is not None):
+        warn_output = "WARNING: You've uploaded an audio file and used the microphone. The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
+        wav = file_mic
+    elif (file_mic is None) and (file_upload is None):
+        return "ERROR: You have to either use the microphone or upload an audio file"
+    elif file_mic is not None:
+        wav = file_mic
+    else:
+        wav = file_upload
+    sig, sr = torchaudio.load(wav)
+    tensor_wav = sig.to(device)
+    resampled = torchaudio.functional.resample( tensor_wav, sr, 16000)
+    sentence = asr.treat_wav(resampled)
+    return sentence
+
+gr.Interface(
+    fn=treat_wav_file,
+    inputs=[gr.Audio(source="microphone", type='filepath', label = "record", optional = True),
+            gr.Audio(source="upload", type='filepath', label="filein", optional=True)]
+    ,outputs="text").launch()
+

TunisianASR/results/14epoch_tunisian/1234/env.log

@@ -5,343 +5,475 @@ Python version:
 [GCC 7.3.0]
 ==============================
 Installed Python packages:
-absl-py==1.2.0
-aiohttp==3.8.1
+abkhazia==1.0
+absl-py==0.11.0
+aiofiles==23.2.1
+aiohttp==3.8.0
 aiosignal==1.2.0
 alabaster==0.7.12
-anaconda-client==1.7.2
-anaconda-navigator==1.10.0
-anaconda-project==0.8.3
+alembic==1.7.4
+altair==4.2.0
+altgraph==0.17
 antlr4-python3-runtime==4.9.3
+anyio==3.6.2
 appdirs==1.4.4
-argh==0.26.2
-argon2-cffi @ file:///tmp/build/80754af9/argon2-cffi_1596828493937/work
-asn1crypto @ file:///tmp/build/80754af9/asn1crypto_1596577642040/work
-astroid @ file:///tmp/build/80754af9/astroid_1592495912941/work
-astropy==4.0.2
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+arrow==1.2.3
+asgiref==3.6.0
+asteroid-filterbanks==0.4.0
+astunparse==1.6.3
 async-generator==1.10
-async-timeout==4.0.2
-atomicwrites==1.4.0
-attrs @ file:///tmp/build/80754af9/attrs_1604765588209/work
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
 audioread==2.1.9
-autopep8 @ file:///tmp/build/80754af9/autopep8_1596578164842/work
-Babel @ file:///tmp/build/80754af9/babel_1605108370292/work
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
 backcall==0.2.0
-backports.functools-lru-cache==1.6.1
-backports.shutil-get-terminal-size==1.0.0
-backports.tempfile==1.0
-backports.weakref==1.0.post1
-beautifulsoup4 @ file:///tmp/build/80754af9/beautifulsoup4_1601924105527/work
-bitarray @ file:///tmp/build/80754af9/bitarray_1605065113847/work
-bkcharts==0.2
-black==22.12.0
-bleach @ file:///tmp/build/80754af9/bleach_1600439572647/work
-bokeh @ file:///tmp/build/80754af9/bokeh_1603297833684/work
-boto==2.49.0
-boto3==1.28.43
-botocore==1.31.43
-Bottleneck==1.3.2
+backports.cached-property==1.0.2
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+blessed==1.20.0
+boto3==1.20.2
+botocore==1.23.2
 bpemb==0.3.4
-brotlipy==0.7.0
-cachetools==5.2.0
-certifi==2020.6.20
-cffi @ file:///tmp/build/80754af9/cffi_1600699146221/work
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
 chardet==3.0.4
-charset-normalizer==2.0.12
-click==8.1.3
-cloudpickle @ file:///tmp/build/80754af9/cloudpickle_1598884132938/work
-clyent==1.2.2
-colorama @ file:///tmp/build/80754af9/colorama_1603211150991/work
-coloredlogs==15.0.1
-conda==4.9.2
-conda-build==3.20.5
-conda-package-handling @ file:///tmp/build/80754af9/conda-package-handling_1603018141399/work
-conda-verify==3.4.2
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cloudpickle==2.2.1
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
 conllu==4.5.3
-contextlib2==0.6.0.post1
-cryptography @ file:///tmp/build/80754af9/cryptography_1601046815590/work
+croniter==1.3.15
+cryptography==38.0.4
+csrgraph==0.1.28
+csvw==1.8.1
 cycler==0.10.0
-Cython @ file:///tmp/build/80754af9/cython_1594831566883/work
-cytoolz==0.11.0
-dask @ file:///tmp/build/80754af9/dask-core_1602083700509/work
-datasets==1.18.3
+Cython==0.29.21
+dataclasses==0.6
+dateutils==0.6.12
 decorator==4.4.2
-defusedxml==0.6.0
+deepdiff==6.3.0
+deepspeech==0.9.1
+defusedxml==0.7.1
 Deprecated==1.2.14
-diff-match-patch @ file:///tmp/build/80754af9/diff-match-patch_1594828741838/work
-dill==0.3.4
-distributed @ file:///tmp/build/80754af9/distributed_1605066520644/work
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
 docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
 easyocr==1.2.1
-einops==0.3.0
+editdistance==0.6.0
+einops==0.3.2
+emoji==2.2.0
 entrypoints==0.3
-et-xmlfile==1.0.1
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
 farasapy==0.0.14
-fastcache==1.1.0
+fastapi==0.98.0
+fastjsonschema==2.17.1
+fasttext==0.9.2
 ffmpeg-python==0.2.0
+ffmpy==0.3.0
 filelock==3.0.12
 flair==0.12.2
-flake8 @ file:///tmp/build/80754af9/flake8_1601911421857/work
-Flask==1.1.2
-flatbuffers==22.9.24
-frozenlist==1.3.0
-fsspec==2022.3.0
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
 ftfy==6.1.1
 future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
 gdown==4.4.0
-gensim==4.1.2
-gevent @ file:///tmp/build/80754af9/gevent_1601397537062/work
-glob2==0.7
-gmpy2==2.0.8
-google-auth==2.12.0
-google-auth-oauthlib==0.4.6
-greenlet @ file:///tmp/build/80754af9/greenlet_1600874013538/work
-grpcio==1.49.1
+gdrive==0.1.5
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-api-core==2.11.1
+google-api-python-client==2.43.0
+google-auth==1.24.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.5.3
+google-pasta==0.2.0
+googleapis-common-protos==1.59.1
+gradio==3.44.4
+gradio-client==0.5.1
+greenlet==1.1.2
+grpcio==1.32.0
+h11==0.14.0
+h5features==1.3.2
 h5py==2.10.0
-HeapDict==1.0.1
-html5lib @ file:///tmp/build/80754af9/html5lib_1593446221756/work
-huggingface-hub==0.16.4
-humanfriendly==10.0
+hierarchy==0.4.0
+hmmlearn==0.2.8
+htk-io==0.5
+httpcore==0.16.3
+httplib2==0.22.0
+httpx==0.23.3
+huggingface-hub==0.15.1
+hydra-colorlog==0.1.4
+hydra-core==1.3.2
 hyperopt==0.2.7
-idna @ file:///tmp/build/80754af9/idna_1593446292537/work
-imageio @ file:///tmp/build/80754af9/imageio_1594161405741/work
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
 imagesize==1.2.0
-imhist==0.0.4
-importlib-metadata==5.0.0
-imWatermark==0.0.2
-iniconfig @ file:///tmp/build/80754af9/iniconfig_1602780191262/work
-install==1.3.5
-intervaltree @ file:///tmp/build/80754af9/intervaltree_1598376443606/work
-invisible-watermark==0.1.5
-ipykernel @ file:///tmp/build/80754af9/ipykernel_1596207638929/work/dist/ipykernel-5.3.4-py3-none-any.whl
-ipython @ file:///tmp/build/80754af9/ipython_1604101197014/work
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+inquirer==3.1.3
+ipadic==1.0.0
+ipyevents==2.0.1
+ipykernel==5.3.4
+ipython==7.19.0
 ipython-genutils==0.2.0
-ipywidgets @ file:///tmp/build/80754af9/ipywidgets_1601490159889/work
-isort @ file:///tmp/build/80754af9/isort_1602603989581/work
-itsdangerous==1.1.0
+ipywebrtc==0.6.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+itsdangerous==2.1.2
 Janome==0.5.0
-jdcal==1.4.1
-jedi @ file:///tmp/build/80754af9/jedi_1592841866100/work
-jeepney @ file:///tmp/build/80754af9/jeepney_1605069705079/work
-Jinja2==2.11.2
-jiwer==2.3.0
-jmespath==1.0.1
-joblib @ file:///tmp/build/80754af9/joblib_1601912903842/work
-json5==0.9.5
-jsonschema @ file:///tmp/build/80754af9/jsonschema_1602607155483/work
-jupyter==1.0.0
-jupyter-client @ file:///tmp/build/80754af9/jupyter_client_1601311786391/work
-jupyter-console @ file:///tmp/build/80754af9/jupyter_console_1598884538475/work
-jupyter-core==4.6.3
-jupyterlab==2.2.6
-jupyterlab-pygments @ file:///tmp/build/80754af9/jupyterlab_pygments_1601490720602/work
-jupyterlab-server @ file:///tmp/build/80754af9/jupyterlab_server_1594164409481/work
-keyring @ file:///tmp/build/80754af9/keyring_1601490835422/work
-kiwisolver @ file:///tmp/build/80754af9/kiwisolver_1604014535162/work
+jedi==0.17.2
+jeepney==0.8.0
+jieba==0.42.1
+Jinja2==3.0.3
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
 langdetect==1.0.9
-lazy-object-proxy==1.4.3
-libarchive-c==2.9
-librosa==0.9.1
-llvmlite==0.34.0
-locket==0.2.0
-lxml @ file:///tmp/build/80754af9/lxml_1603216285000/work
-Markdown==3.4.1
-MarkupSafe==1.1.1
-matplotlib @ file:///tmp/build/80754af9/matplotlib-base_1603378225747/work
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+lightning-cloud==0.5.37
+lightning-utilities==0.8.0
+linkify-it-py==1.0.3
+lit==16.0.6
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+markdown-it-py==3.0.0
+MarkupSafe==2.1.3
+marshmallow==3.14.0
+matplotlib==3.3.3
 mccabe==0.6.1
+mcd==0.4
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+metrics==0.3.3
 mido==1.2.10
 mistune==0.8.4
-mkl-fft==1.2.0
-mkl-random==1.1.1
-mkl-service==2.3.0
-mock==4.0.2
-more-itertools @ file:///tmp/build/80754af9/more-itertools_1605111547926/work
+more-itertools==8.6.0
 mpld3==0.3
-mpmath==1.1.0
-msgpack==1.0.0
-multidict==6.0.2
-multipledispatch==0.6.0
-multiprocess==0.70.12.2
-mypy-extensions==0.4.3
-navigator-updater==0.2.1
-nbclient @ file:///tmp/build/80754af9/nbclient_1602783176460/work
-nbconvert @ file:///tmp/build/80754af9/nbconvert_1601914830498/work
-nbformat @ file:///tmp/build/80754af9/nbformat_1602783287752/work
-nest-asyncio @ file:///tmp/build/80754af9/nest-asyncio_1605115881283/work
-networkx @ file:///tmp/build/80754af9/networkx_1598376031484/work
-nltk @ file:///tmp/build/80754af9/nltk_1592496090529/work
-nose==1.3.7
-notebook @ file:///tmp/build/80754af9/notebook_1601501575118/work
-numba @ file:///tmp/build/80754af9/numba_1600100669015/work
-numexpr==2.7.1
-numpy @ file:///tmp/build/80754af9/numpy_and_numpy_base_1603570489231/work
-numpydoc @ file:///tmp/build/80754af9/numpydoc_1605117425582/work
-oauthlib==3.2.1
-olefile==0.46
-omegaconf==2.2.3
-onnx==1.12.0
-onnxruntime==1.12.1
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==5.6.1
+nbformat==5.9.0
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.8.8
+nltk==3.2.4
+nodeenv==1.5.0
+normalize==2.0.2
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-cupti-cu11==11.7.101
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+nvidia-cufft-cu11==10.9.0.58
+nvidia-curand-cu11==10.2.10.91
+nvidia-cusolver-cu11==11.4.0.1
+nvidia-cusparse-cu11==11.7.4.91
+nvidia-nccl-cu11==2.14.3
+nvidia-nvtx-cu11==11.7.91
+oauthlib==3.1.0
+omegaconf==2.3.0
+onnx==1.10.2
+OpenCC==1.1.2
 opencv-python==4.4.0.46
-openpyxl @ file:///tmp/build/80754af9/openpyxl_1598113097404/work
-packaging==20.9
-pandas @ file:///tmp/build/80754af9/pandas_1602088120436/work
-pandocfilters @ file:///tmp/build/80754af9/pandocfilters_1605120460739/work
-parso==0.7.0
-partd==1.1.0
-path @ file:///tmp/build/80754af9/path_1598376507494/work
-pathlib2==2.3.5
-pathspec==0.10.3
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+ordered-set==4.1.0
+orjson==3.8.4
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathlib2==2.3.7.post1
+pathspec==0.5.5
 pathtools==0.1.2
-patsy==0.5.1
-pep8==1.7.1
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
 pexpect==4.8.0
+phonemizer==2.2.1
 pickleshare==0.7.5
-Pillow @ file:///tmp/build/80754af9/pillow_1603822255246/work
-pkginfo==1.6.1
-platformdirs==2.6.0
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
 pluggy==0.13.1
-ply==3.11
-pooch==1.6.0
+pooch==1.3.0
+portalocker==2.3.2
 pptree==3.1
+pre-commit==2.9.0
+preprocessing==0.1.13
 pretty-midi==0.2.9
-prometheus-client==0.8.0
-prompt-toolkit @ file:///tmp/build/80754af9/prompt-toolkit_1602688806899/work
-protobuf==3.19.6
-psutil @ file:///tmp/build/80754af9/psutil_1598370257551/work
+prettytable==2.2.1
+primePy==1.3
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.20.3
+psutil==5.6.6
 ptyprocess==0.6.0
-py @ file:///tmp/build/80754af9/py_1593446248552/work
+py==1.9.0
 py-espeak-ng==0.1.8
 py4j==0.10.9.7
+pyannote.audio==2.1.1
+pyannote.core==4.5
+pyannote.database==4.1.3
+pyannote.metrics==3.2.1
+pyannote.pipeline==2.3
+pyannotebook==0.1.0.dev0
 PyArabic==0.6.15
-pyarrow==7.0.0
+pyarrow==3.0.0
 pyasn1==0.4.8
 pyasn1-modules==0.2.8
-pycodestyle==2.6.0
-pycosat==0.6.3
-pycparser @ file:///tmp/build/80754af9/pycparser_1594388511720/work
-pycurl==7.43.0.6
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pycryptodome==3.16.0
+pyctcdecode==0.4.0
+pydantic==1.10.4
 pyDeprecate==0.3.1
-pydocstyle @ file:///tmp/build/80754af9/pydocstyle_1598885001695/work
-pyflakes==2.2.0
-Pygments @ file:///tmp/build/80754af9/pygments_1604103097372/work
-pylint @ file:///tmp/build/80754af9/pylint_1598623985952/work
-pyodbc===4.0.0-unsupported
-pyOpenSSL @ file:///tmp/build/80754af9/pyopenssl_1594392929924/work
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.15.1
+pygtrie==2.5.0
+PyJWT==2.7.0
+pymodbus==2.5.3
 pyparsing==2.4.7
-pyrsistent @ file:///tmp/build/80754af9/pyrsistent_1600141720057/work
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
 PySocks==1.7.1
-pytest==0.0.0
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
 python-bidi==0.4.2
 python-crfsuite==0.9.7
-python-dateutil==2.8.1
-python-jsonrpc-server @ file:///tmp/build/80754af9/python-jsonrpc-server_1600278539111/work
-python-language-server @ file:///tmp/build/80754af9/python-language-server_1600454544709/work
+python-dateutil==2.8.2
+python-editor==1.0.4
 python-Levenshtein==0.12.2
-pytorch-lightning==1.4.2
+python-multipart==0.0.5
+python-utils==2.4.0
+pytorch-lightning==1.6.5
+pytorch-metric-learning==1.7.3
 pytorch-revgrad==0.2.0
-pytz==2020.1
-PyWavelets @ file:///tmp/build/80754af9/pywavelets_1601658317819/work
-pyxdg @ file:///tmp/build/80754af9/pyxdg_1603822279816/work
-PyYAML==5.3.1
-pyzmq==19.0.2
-QDarkStyle==2.8.1
-QtAwesome @ file:///tmp/build/80754af9/qtawesome_1602272867890/work
-qtconsole @ file:///tmp/build/80754af9/qtconsole_1600870028330/work
-QtPy==1.9.0
-regex @ file:///tmp/build/80754af9/regex_1602786672676/work
-requests @ file:///tmp/build/80754af9/requests_1592841827918/work
-requests-oauthlib==1.3.1
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==6.0
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+readchar==4.0.5
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
 resampy==0.2.2
-rope @ file:///tmp/build/80754af9/rope_1602264064449/work
-rsa==4.9
-Rtree==0.9.4
-ruamel-yaml==0.15.87
-s3transfer==0.6.2
-sacremoses==0.0.49
-safetensors==0.3.3
-scikit-image==0.17.2
-scikit-learn @ file:///tmp/build/80754af9/scikit-learn_1598376899566/work
-scipy @ file:///tmp/build/80754af9/scipy_1597686649129/work
-seaborn @ file:///tmp/build/80754af9/seaborn_1600553570093/work
-SecretStorage==3.1.2
+rfc3986==1.4.0
+rich==13.4.2
+richenum==1.3.1
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+safetensors==0.3.1
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+SecretStorage==3.3.3
+segments==2.1.3
 segtok==1.5.11
+semantic-version==2.10.0
+semver==2.13.0
 Send2Trash==1.5.0
-sentencepiece==0.1.97
-simplegeneric==0.8.1
-singledispatch @ file:///tmp/build/80754af9/singledispatch_1602523705405/work
-sip==4.19.13
-six @ file:///tmp/build/80754af9/six_1605205327372/work
-smart-open==5.2.1
+sentencepiece==0.1.99
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+singledispatchmethod==1.0
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+sniffio==1.3.0
 snowballstemmer==2.0.0
-sortedcollections==1.2.1
-sortedcontainers==2.2.2
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
 SoundFile==0.10.3.post1
-soupsieve==2.0.1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.14
 sphfile==1.0.3
-Sphinx @ file:///tmp/build/80754af9/sphinx_1597428793432/work
+Sphinx==3.3.1
+sphinx-rtd-theme==0.2.4
 sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
 sphinxcontrib-devhelp==1.0.2
 sphinxcontrib-htmlhelp==1.0.3
 sphinxcontrib-jsmath==1.0.1
 sphinxcontrib-qthelp==1.0.3
 sphinxcontrib-serializinghtml==1.1.4
-sphinxcontrib-websupport @ file:///tmp/build/80754af9/sphinxcontrib-websupport_1597081412696/work
-spyder @ file:///tmp/build/80754af9/spyder_1599056981321/work
-spyder-kernels @ file:///tmp/build/80754af9/spyder-kernels_1599056754858/work
-SQLAlchemy @ file:///tmp/build/80754af9/sqlalchemy_1603397987316/work
+SQLAlchemy==1.4.25
 sqlitedict==2.1.0
-statsmodels @ file:///tmp/build/80754af9/statsmodels_1602280205159/work
-sympy @ file:///tmp/build/80754af9/sympy_1605119542615/work
-tables==3.6.1
-tabulate==0.9.0
-tblib @ file:///tmp/build/80754af9/tblib_1597928476713/work
-tensorboard==2.10.1
-tensorboard-data-server==0.6.1
-tensorboard-plugin-wit==1.8.1
-terminado==0.9.1
+sqlparse==0.4.2
+stanza==1.4.2
+starlette==0.27.0
+starsessions==1.3.0
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorboardX==2.6.1
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
 testpath==0.4.4
-threadpoolctl @ file:///tmp/tmp9twdgx9k/threadpoolctl-2.1.0-py3-none-any.whl
-tifffile==2020.10.1
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tinycss2==1.2.1
 tkseem==0.0.3
 tokenizers==0.13.3
-toml @ file:///tmp/build/80754af9/toml_1592853716807/work
-tomli==2.0.1
-toolz @ file:///tmp/build/80754af9/toolz_1601054250827/work
-torch==1.11.0
-torchaudio==0.11.0
-torchmetrics==0.6.0
-torchvision==0.8.2
-tornado==6.0.4
-tqdm==4.64.0
-traitlets @ file:///tmp/build/80754af9/traitlets_1602787416690/work
+toml==0.10.2
+toolz==0.12.0
+torch==1.13.1
+torch-audiomentations==0.11.0
+torch-pitch-shift==1.2.4
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.11.4
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.9.0
 transformer-smaller-training-vocab==0.3.1
-transformers==4.33.1
+transformers==4.30.2
+triton==2.0.0
+typed-ast==1.4.1
+typer==0.4.0
 typing-extensions==4.4.0
-ujson @ file:///tmp/build/80754af9/ujson_1602523317881/work
-unicodecsv==0.14.1
-urllib3 @ file:///tmp/build/80754af9/urllib3_1603305693037/work
-watchdog @ file:///tmp/build/80754af9/watchdog_1593447344699/work
-wcwidth @ file:///tmp/build/80754af9/wcwidth_1593447189090/work
+uc-micro-py==1.0.1
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+uvicorn==0.20.0
+versioneer==0.28
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
 webencodings==0.5.1
+websocket-client==1.6.1
+websockets==10.4
 Werkzeug==1.0.1
+wget==3.2
 widgetsnbextension==3.5.1
 Wikipedia-API==0.6.0
-wrapt==1.11.2
-wurlitzer @ file:///tmp/build/80754af9/wurlitzer_1594753850195/work
-xlrd==1.2.0
-XlsxWriter @ file:///tmp/build/80754af9/xlsxwriter_1602692860603/work
-xlwt==1.3.0
-xmltodict==0.12.0
-xxhash==3.0.0
-yapf @ file:///tmp/build/80754af9/yapf_1593528177422/work
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
 yarl==1.7.2
-zict==2.0.0
-zipp @ file:///tmp/build/80754af9/zipp_1604001098328/work
-zope.event==4.5.0
-zope.interface @ file:///tmp/build/80754af9/zope.interface_1602002420968/work
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
 ==============================
 Git revision:
-8a51838
+be9098b
 ==============================
 CUDA version:
 11.7

TunisianASR/results/14epoch_tunisian/1234/hyperparams.yaml

@@ -1,5 +1,5 @@
-# Generated 2023-09-20 from:
-# /home/salah/Code_Switched_Tunisian_Speech_Recognition/TunisianASR/semi_trained.yaml
+# Generated 2023-09-25 from:
+# /home/salah/Code-Switched-Tunisian-SpeechToText/TunisianASR/semi_trained.yaml
 # yamllint disable
 # ################################
 # Model: wav2vec2 + DNN + CTC

TunisianASR/results/14epoch_tunisian/1234/log.txt

@@ -357,3 +357,494 @@ zope.interface @ file:///tmp/build/80754af9/zope.interface_1602002420968/work
 2023-09-20 16:24:00,139 - speechbrain.core - INFO - 314.4M trainable parameters in ASR
 2023-09-20 16:24:00,967 - speechbrain.utils.checkpoints - INFO - Loading a checkpoint from TunisianASR/results/14epoch_tunisian/1234/save/CKPT+2023-08-03+01-38-38+00
 2023-09-20 16:24:49,007 - speechbrain.utils.distributed - INFO - distributed_launch flag is disabled, this experiment will be executed without DDP.
+2023-09-25 11:12:54,556 - speechbrain.core - INFO - Beginning experiment!
+2023-09-25 11:12:54,556 - speechbrain.core - INFO - Experiment folder: TunisianASR/results/14epoch_tunisian/1234/
+2023-09-25 11:12:55,141 - speechbrain.utils.superpowers - DEBUG - abkhazia==1.0
+absl-py==0.11.0
+aiofiles==23.2.1
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altair==4.2.0
+altgraph==0.17
+antlr4-python3-runtime==4.9.3
+anyio==3.6.2
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+arrow==1.2.3
+asgiref==3.6.0
+asteroid-filterbanks==0.4.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+backports.cached-property==1.0.2
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+blessed==1.20.0
+boto3==1.20.2
+botocore==1.23.2
+bpemb==0.3.4
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cloudpickle==2.2.1
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+conllu==4.5.3
+croniter==1.3.15
+cryptography==38.0.4
+csrgraph==0.1.28
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+dateutils==0.6.12
+decorator==4.4.2
+deepdiff==6.3.0
+deepspeech==0.9.1
+defusedxml==0.7.1
+Deprecated==1.2.14
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+einops==0.3.2
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fastapi==0.98.0
+fastjsonschema==2.17.1
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+ffmpy==0.3.0
+filelock==3.0.12
+flair==0.12.2
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+ftfy==6.1.1
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.4.0
+gdrive==0.1.5
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-api-core==2.11.1
+google-api-python-client==2.43.0
+google-auth==1.24.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.5.3
+google-pasta==0.2.0
+googleapis-common-protos==1.59.1
+gradio==3.44.4
+gradio-client==0.5.1
+greenlet==1.1.2
+grpcio==1.32.0
+h11==0.14.0
+h5features==1.3.2
+h5py==2.10.0
+hierarchy==0.4.0
+hmmlearn==0.2.8
+htk-io==0.5
+httpcore==0.16.3
+httplib2==0.22.0
+httpx==0.23.3
+huggingface-hub==0.15.1
+hydra-colorlog==0.1.4
+hydra-core==1.3.2
+hyperopt==0.2.7
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+inquirer==3.1.3
+ipadic==1.0.0
+ipyevents==2.0.1
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywebrtc==0.6.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+itsdangerous==2.1.2
+Janome==0.5.0
+jedi==0.17.2
+jeepney==0.8.0
+jieba==0.42.1
+Jinja2==3.0.3
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+langdetect==1.0.9
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+lightning-cloud==0.5.37
+lightning-utilities==0.8.0
+linkify-it-py==1.0.3
+lit==16.0.6
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+markdown-it-py==3.0.0
+MarkupSafe==2.1.3
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+metrics==0.3.3
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpld3==0.3
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==5.6.1
+nbformat==5.9.0
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.8.8
+nltk==3.2.4
+nodeenv==1.5.0
+normalize==2.0.2
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-cupti-cu11==11.7.101
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+nvidia-cufft-cu11==10.9.0.58
+nvidia-curand-cu11==10.2.10.91
+nvidia-cusolver-cu11==11.4.0.1
+nvidia-cusparse-cu11==11.7.4.91
+nvidia-nccl-cu11==2.14.3
+nvidia-nvtx-cu11==11.7.91
+oauthlib==3.1.0
+omegaconf==2.3.0
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+ordered-set==4.1.0
+orjson==3.8.4
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathlib2==2.3.7.post1
+pathspec==0.5.5
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pptree==3.1
+pre-commit==2.9.0
+preprocessing==0.1.13
+pretty-midi==0.2.9
+prettytable==2.2.1
+primePy==1.3
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.20.3
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+py4j==0.10.9.7
+pyannote.audio==2.1.1
+pyannote.core==4.5
+pyannote.database==4.1.3
+pyannote.metrics==3.2.1
+pyannote.pipeline==2.3
+pyannotebook==0.1.0.dev0
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pycryptodome==3.16.0
+pyctcdecode==0.4.0
+pydantic==1.10.4
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.15.1
+pygtrie==2.5.0
+PyJWT==2.7.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-editor==1.0.4
+python-Levenshtein==0.12.2
+python-multipart==0.0.5
+python-utils==2.4.0
+pytorch-lightning==1.6.5
+pytorch-metric-learning==1.7.3
+pytorch-revgrad==0.2.0
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==6.0
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+readchar==4.0.5
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rich==13.4.2
+richenum==1.3.1
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+safetensors==0.3.1
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+SecretStorage==3.3.3
+segments==2.1.3
+segtok==1.5.11
+semantic-version==2.10.0
+semver==2.13.0
+Send2Trash==1.5.0
+sentencepiece==0.1.99
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+singledispatchmethod==1.0
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+sniffio==1.3.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.14
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.2.4
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlitedict==2.1.0
+sqlparse==0.4.2
+stanza==1.4.2
+starlette==0.27.0
+starsessions==1.3.0
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorboardX==2.6.1
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tinycss2==1.2.1
+tkseem==0.0.3
+tokenizers==0.13.3
+toml==0.10.2
+toolz==0.12.0
+torch==1.13.1
+torch-audiomentations==0.11.0
+torch-pitch-shift==1.2.4
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.11.4
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.9.0
+transformer-smaller-training-vocab==0.3.1
+transformers==4.30.2
+triton==2.0.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==4.4.0
+uc-micro-py==1.0.1
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+uvicorn==0.20.0
+versioneer==0.28
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+websocket-client==1.6.1
+websockets==10.4
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+Wikipedia-API==0.6.0
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+
+
+2023-09-25 11:12:55,173 - speechbrain.utils.superpowers - DEBUG - be9098b
+
+
+2023-09-25 11:12:55,216 - speechbrain.pretrained.fetching - INFO - Fetch hyperparams.yaml: Using existing file/symlink in pretrained_models/asr-wav2vec2-commonvoice-fr/hyperparams.yaml.
+2023-09-25 11:12:55,217 - speechbrain.pretrained.fetching - INFO - Fetch custom.py: Linking to local file in /home/salah/Code-Switched-Tunisian-SpeechToText/asr-wav2vec2-commonvoice-fr/custom.py.
+2023-09-25 11:12:58,078 - speechbrain.lobes.models.huggingface_wav2vec - WARNING - speechbrain.lobes.models.huggingface_wav2vec - wav2vec 2.0 is frozen.
+2023-09-25 11:12:58,080 - speechbrain.utils.parameter_transfer - DEBUG - Collecting files (or symlinks) for pretraining in pretrained_models/asr-wav2vec2-commonvoice-fr.
+2023-09-25 11:12:58,087 - speechbrain.pretrained.fetching - INFO - Fetch wav2vec2.ckpt: Using existing file/symlink in pretrained_models/asr-wav2vec2-commonvoice-fr/wav2vec2.ckpt.
+2023-09-25 11:12:58,087 - speechbrain.pretrained.fetching - INFO - Fetch asr.ckpt: Using existing file/symlink in pretrained_models/asr-wav2vec2-commonvoice-fr/asr.ckpt.
+2023-09-25 11:12:58,087 - speechbrain.pretrained.fetching - INFO - Fetch tokenizer.ckpt: Using existing file/symlink in pretrained_models/asr-wav2vec2-commonvoice-fr/tokenizer.ckpt.
+2023-09-25 11:12:58,087 - speechbrain.utils.parameter_transfer - INFO - Loading pretrained files for: wav2vec2, asr, tokenizer
+2023-09-25 11:13:01,875 - speechbrain.lobes.models.huggingface_wav2vec - WARNING - speechbrain.lobes.models.huggingface_wav2vec - wav2vec 2.0 feature extractor is frozen.
+2023-09-25 11:13:01,877 - speechbrain.core - INFO - Info: auto_mix_prec arg from hparam file is used
+2023-09-25 11:13:01,877 - speechbrain.core - INFO - Info: ckpt_interval_minutes arg from hparam file is used
+2023-09-25 11:13:01,880 - speechbrain.core - INFO - 314.4M trainable parameters in ASRCV
+2023-09-25 11:13:01,885 - speechbrain.utils.checkpoints - INFO - Loading a checkpoint from EnglishCV/results/wav2vec2_ctc_en/1234/save/CKPT+2023-09-06+22-56-31+00
+2023-09-25 11:13:04,505 - speechbrain.core - INFO - Info: auto_mix_prec arg from hparam file is used
+2023-09-25 11:13:04,505 - speechbrain.core - INFO - Info: ckpt_interval_minutes arg from hparam file is used
+2023-09-25 11:13:04,509 - speechbrain.core - INFO - 314.4M trainable parameters in ASR
+2023-09-25 11:13:04,513 - speechbrain.utils.checkpoints - INFO - Loading a checkpoint from TunisianASR/results/14epoch_tunisian/1234/save/CKPT+2023-08-03+01-38-38+00
+2023-09-25 11:13:05,900 - speechbrain.utils.distributed - INFO - distributed_launch flag is disabled, this experiment will be executed without DDP.

app.py

@@ -356,7 +356,7 @@ english_asr_model = ASRCV(
     )
 english_asr_model.modules.to("cpu")
 english_asr_model.device="cpu"
-#english_asr_model.checkpointer.recover_if_possible()
+english_asr_model.checkpointer.recover_if_possible(device="cpu")
 run_opts["device"]="cpu"
 print("moving to tunisian model")
 asr_brain = ASR(
@@ -366,7 +366,7 @@ asr_brain = ASR(
     checkpointer=hparams["checkpointer"],
 )
 asr_brain.modules.to("cpu")
-#asr_brain.checkpointer.recover_if_possible()
+asr_brain.checkpointer.recover_if_possible(device="cpu")
 asr_brain.modules.eval()
 english_asr_model.modules.eval()
 french_asr_model.mods.eval()
@@ -713,7 +713,7 @@ mixer = Mixer(
 )
 mixer.tokenizer = label_encoder
 mixer.device = "cpu"
-#mixer.checkpointer.recover_if_possible()
+mixer.checkpointer.recover_if_possible(device="cpu")
 mixer.modules.eval()
 
 

pretrained_models/asr-wav2vec2-commonvoice-fr/custom.py

@@ -1 +1 @@
-/home/salah/Code_Switched_Tunisian_Speech_Recognition/asr-wav2vec2-commonvoice-fr/custom.py
+/home/salah/Code-Switched-Tunisian-SpeechToText/asr-wav2vec2-commonvoice-fr/custom.py

results/non_semi_final_stac/app.py

@@ -0,0 +1,772 @@
+import os
+import sys
+import torch
+import logging
+import speechbrain as sb
+from speechbrain.utils.distributed import run_on_main
+from hyperpyyaml import load_hyperpyyaml
+from pathlib import Path
+import torchaudio.transforms as T
+from cv_train import ASRCV
+import torchaudio
+import numpy as np
+import kenlm
+from pyctcdecode import build_ctcdecoder
+import re
+from torch.nn.utils.rnn import pad_sequence
+import torch.optim as optim
+import torch.nn as nn
+
+
+# Commented out IPython magic to ensure Python compatibility.
+hparams_file, run_opts, overrides = sb.parse_arguments(["TunisianASR/semi_trained.yaml"])
+
+# If distributed_launch=True then
+# create ddp_group with the right communication protocol
+sb.utils.distributed.ddp_init_group(run_opts)
+
+with open(hparams_file) as fin:
+    hparams = load_hyperpyyaml(fin, overrides)
+
+# Create experiment directory
+sb.create_experiment_directory(
+    experiment_directory=hparams["output_folder"],
+    hyperparams_to_save=hparams_file,
+    overrides=overrides,
+)
+# Dataset prep (parsing Librispeech)
+
+def dataio_prepare(hparams):
+    """This function prepares the datasets to be used in the brain class.
+    It also defines the data processing pipeline through user-defined functions."""
+
+    # 1. Define datasets
+    data_folder = hparams["data_folder"]
+
+    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
+    )
+
+    if hparams["sorting"] == "ascending":
+        # we sort training data to speed up training and get better results.
+        train_data = train_data.filtered_sorted(
+            sort_key="duration",
+            key_max_value={"duration": hparams["avoid_if_longer_than"]},
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["dataloader_options"]["shuffle"] = False
+
+    elif hparams["sorting"] == "descending":
+        train_data = train_data.filtered_sorted(
+            sort_key="duration",
+            reverse=True,
+            key_max_value={"duration": hparams["avoid_if_longer_than"]},
+        )
+        # when sorting do not shuffle in dataloader ! otherwise is pointless
+        hparams["dataloader_options"]["shuffle"] = False
+
+    elif hparams["sorting"] == "random":
+        pass
+
+    else:
+        raise NotImplementedError(
+            "sorting must be random, ascending or descending"
+        )
+
+    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
+        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
+    )
+    # We also sort the validation data so it is faster to validate
+    valid_data = valid_data.filtered_sorted(sort_key="duration")
+    test_datasets = {}
+    for csv_file in hparams["test_csv"]:
+        name = Path(csv_file).stem
+        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
+            csv_path=csv_file, replacements={"data_root": data_folder}
+        )
+        test_datasets[name] = test_datasets[name].filtered_sorted(
+            sort_key="duration"
+        )
+
+    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
+
+
+    # 2. Define audio pipeline:
+    @sb.utils.data_pipeline.takes("wav")
+    @sb.utils.data_pipeline.provides("sig")
+    def audio_pipeline(wav):
+        info = torchaudio.info(wav)
+        sig = sb.dataio.dataio.read_audio(wav)
+        if len(sig.shape)>1 :
+            sig = torch.mean(sig, dim=1)
+        resampled = torchaudio.transforms.Resample(
+            info.sample_rate, hparams["sample_rate"],
+        )(sig)
+        return resampled
+
+    sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
+    label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+    # 3. Define text pipeline:
+    @sb.utils.data_pipeline.takes("wrd")
+    @sb.utils.data_pipeline.provides(
+        "wrd", "char_list", "tokens_list", "tokens"
+    )
+    def text_pipeline(wrd):
+        yield wrd
+        char_list = list(wrd)
+        yield char_list
+        tokens_list = label_encoder.encode_sequence(char_list)
+        yield tokens_list
+        tokens = torch.LongTensor(tokens_list)
+        yield tokens
+
+    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
+    lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+    special_labels = {
+        "blank_label": hparams["blank_index"],
+        "unk_label": hparams["unk_index"]
+    }
+    label_encoder.load_or_create(
+        path=lab_enc_file,
+        from_didatasets=[train_data],
+        output_key="char_list",
+        special_labels=special_labels,
+        sequence_input=True,
+    )
+
+    # 4. Set output:
+    sb.dataio.dataset.set_output_keys(
+        datasets, ["id", "sig", "wrd", "char_list", "tokens"],
+    )
+    return train_data, valid_data,test_datasets, label_encoder
+
+class ASR(sb.core.Brain):
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+
+        batch = batch.to(self.device)
+        wavs, wav_lens = batch.sig
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        if stage == sb.Stage.TRAIN:
+            if hasattr(self.hparams, "augmentation"):
+                wavs = self.hparams.augmentation(wavs, wav_lens)
+
+        # Forward pass
+        feats = self.modules.wav2vec2(wavs, wav_lens)
+        x = self.modules.enc(feats)
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+
+        return p_ctc, wav_lens
+
+    def custom_encode(self,wavs,wav_lens) :
+        wavs = wavs.to("cpu")
+        if(wav_lens is not None): wav_lens.to(self.device)
+
+        feats = self.modules.wav2vec2(wavs, wav_lens)
+        x = self.modules.enc(feats)
+        logits = self.modules.ctc_lin(x)
+        p_ctc = self.hparams.log_softmax(logits)
+
+        return feats,p_ctc
+
+
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC) given predictions and targets."""
+
+        p_ctc, wav_lens = predictions
+
+        ids = batch.id
+        tokens, tokens_lens = batch.tokens
+
+        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+
+        if stage != sb.Stage.TRAIN:
+            predicted_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+            # Decode token terms to words
+            if self.hparams.use_language_modelling:
+                predicted_words = []
+                for logs in p_ctc:
+                    text = decoder.decode(logs.detach().cpu().numpy())
+                    predicted_words.append(text.split(" "))
+            else:
+                predicted_words = [
+                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                    for utt_seq in predicted_tokens
+                ]
+            # Convert indices to words
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        should_step = self.step % self.grad_accumulation_factor == 0
+        # Managing automatic mixed precision
+        # TOFIX: CTC fine-tuning currently is unstable
+        # This is certainly due to CTC being done in fp16 instead of fp32
+        if self.auto_mix_prec:
+            with torch.cuda.amp.autocast():
+                with self.no_sync():
+                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+            with self.no_sync(not should_step):
+                self.scaler.scale(
+                    loss / self.grad_accumulation_factor
+                ).backward()
+            if should_step:
+
+                if not self.hparams.wav2vec2.freeze:
+                    self.scaler.unscale_(self.wav2vec_optimizer)
+                self.scaler.unscale_(self.model_optimizer)
+                if self.check_gradients(loss):
+                    if not self.hparams.wav2vec2.freeze:
+                        if self.optimizer_step >= self.hparams.warmup_steps:
+                            self.scaler.step(self.wav2vec_optimizer)
+                    self.scaler.step(self.model_optimizer)
+                self.scaler.update()
+                self.zero_grad()
+                self.optimizer_step += 1
+        else:
+            # This is mandatory because HF models have a weird behavior with DDP
+            # on the forward pass
+            with self.no_sync():
+                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+
+            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+
+            with self.no_sync(not should_step):
+                (loss / self.grad_accumulation_factor).backward()
+            if should_step:
+                if self.check_gradients(loss):
+                    if not self.hparams.wav2vec2.freeze:
+                        if self.optimizer_step >= self.hparams.warmup_steps:
+                            self.wav2vec_optimizer.step()
+                    self.model_optimizer.step()
+                self.zero_grad()
+                self.optimizer_step += 1
+
+        self.on_fit_batch_end(batch, outputs, loss, should_step)
+        return loss.detach().cpu()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            if not self.hparams.wav2vec2.freeze:
+                sb.nnet.schedulers.update_learning_rate(
+                    self.wav2vec_optimizer, new_lr_wav2vec
+                )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                    "lr_wav2vec": old_lr_wav2vec,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+        "Initializes the wav2vec2 optimizer and model optimizer"
+
+        # If the wav2vec encoder is unfrozen, we create the optimizer
+        if not self.hparams.wav2vec2.freeze:
+            self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
+                self.modules.wav2vec2.parameters()
+            )
+            if self.checkpointer is not None:
+                self.checkpointer.add_recoverable(
+                    "wav2vec_opt", self.wav2vec_optimizer
+                )
+
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+    def zero_grad(self, set_to_none=False):
+        if not self.hparams.wav2vec2.freeze:
+            self.wav2vec_optimizer.zero_grad(set_to_none)
+        self.model_optimizer.zero_grad(set_to_none)
+
+
+from speechbrain.pretrained import EncoderASR,EncoderDecoderASR
+french_asr_model = EncoderASR.from_hparams(source="asr-wav2vec2-commonvoice-fr", savedir="pretrained_models/asr-wav2vec2-commonvoice-fr")
+french_asr_model.to("cpu")
+cvhparams_file, cvrun_opts, cvoverrides = sb.parse_arguments(["EnglishCV/train_en_with_wav2vec.yaml"])
+with open(cvhparams_file) as cvfin:
+    cvhparams = load_hyperpyyaml(cvfin, cvoverrides)
+cvrun_opts["device"]="cpu"
+english_asr_model = ASRCV(
+        modules=cvhparams["modules"],
+        hparams=cvhparams,
+        run_opts=cvrun_opts,
+        checkpointer=cvhparams["checkpointer"],
+    )
+english_asr_model.modules.to("cpu")
+english_asr_model.device="cpu"
+english_asr_model.checkpointer.recover_if_possible()
+run_opts["device"]="cpu"
+print("moving to tunisian model")
+asr_brain = ASR(
+    modules=hparams["modules"],
+    hparams=hparams,
+    run_opts=run_opts,
+    checkpointer=hparams["checkpointer"],
+)
+asr_brain.modules.to("cpu")
+asr_brain.checkpointer.recover_if_possible()
+asr_brain.modules.eval()
+english_asr_model.modules.eval()
+french_asr_model.mods.eval()
+asr_brain.modules.to("cpu")
+
+# Commented out IPython magic to ensure Python compatibility.
+# %ls
+
+#UTILS FUNCTIOJNS
+def get_size_dimensions(arr):
+    size_dimensions = []
+    while isinstance(arr, list):
+        size_dimensions.append(len(arr))
+        arr = arr[0]
+    return size_dimensions
+
+def scale_array(batch,n):
+    scaled_batch = []
+
+    for array in batch:
+        if(n < len(array)): raise ValueError("Cannot scale Array down")
+
+        repeat = round(n/len(array))+1
+        scaled_length_array= []
+
+        for i in array:
+            for j in range(repeat) :
+                if(len(scaled_length_array) == n): break
+                scaled_length_array.append(i)
+
+        scaled_batch.append(scaled_length_array)
+
+    return torch.tensor(scaled_batch)
+
+
+def load_paths(wavs_path):
+    waveforms = []
+    for path in wavs_path :
+        waveform, _ = torchaudio.load(path)
+        waveforms.append(waveform.squeeze(0))
+    # normalize array length to the bigger arrays by pading with 0's
+    padded_arrays = pad_sequence(waveforms, batch_first=True)
+    return torch.tensor(padded_arrays)
+
+
+
+device = 'cpu'
+verbose = 0
+#FLOW LEVEL FUNCTIONS
+def merge_strategy(embeddings1, embeddings2, embeddings3,post1, post2,post3):
+
+
+    post1 = post1.to(device)
+    post2 = post2.to(device)
+    post3 = post3.to(device)
+    embeddings1 = embeddings1.to(device)
+    embeddings2 = embeddings2.to(device)
+    embeddings3 = embeddings3.to(device)
+
+    posteriograms_merged = torch.cat((post1,post2,post3),dim=2)
+    embeddings_merged = torch.cat((embeddings1,embeddings2,embeddings3),dim=2)
+
+    if(verbose !=0):
+      print('MERGED POST ',posteriograms_merged.shape)
+      print('MERGED emb ',embeddings_merged.shape)
+
+    return torch.cat((posteriograms_merged,embeddings_merged),dim=2).to(device)
+
+def decode(model,wavs,wav_lens):
+
+    with torch.no_grad():
+        wav_lens = wav_lens.to(model.device)
+        encoder_out = model.encode_batch(wavs, wav_lens)
+        predictions = model.decoding_function(encoder_out, wav_lens)
+        return predictions
+
+def middle_layer(batch, lens):
+
+    tn_embeddings, tn_posteriogram = asr_brain.custom_encode(batch,None)
+
+    fr_embeddings = french_asr_model.mods.encoder.wav2vec2(batch)
+    fr_posteriogram =french_asr_model.encode_batch(batch,lens)
+    en_embeddings = english_asr_model.modules.wav2vec2(batch, lens)
+    x = english_asr_model.modules.enc(en_embeddings)
+    en_posteriogram = english_asr_model.modules.ctc_lin(x)
+    #scores, en_posteriogram = english_asr_model.mods.decoder(en_embeddings ,lens)
+    if(verbose !=0):
+      print('[EMBEDDINGS] FR:',fr_embeddings.shape, "EN:",en_embeddings.shape, "TN:", tn_embeddings.shape)
+      print('[POSTERIOGRAM] FR:',fr_posteriogram.shape, "EN:",en_posteriogram.shape,"TN:",tn_posteriogram.shape)
+
+
+    bilangual_sample = merge_strategy(fr_embeddings,en_embeddings,tn_embeddings,fr_posteriogram,en_posteriogram,tn_posteriogram)
+    return bilangual_sample
+
+class Mixer(sb.core.Brain):
+
+    def compute_forward(self, batch, stage):
+        """Forward computations from the waveform batches to the output probabilities."""
+        wavs, wav_lens = batch.sig
+        wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
+
+        if stage == sb.Stage.TRAIN:
+            if hasattr(self.hparams, "augmentation"):
+                wavs = self.hparams.augmentation(wavs, wav_lens)
+
+        multi_langual_feats = middle_layer(wavs, wav_lens)
+        multi_langual_feats= multi_langual_feats.to(device)
+        feats, _ = self.modules.enc(multi_langual_feats)
+        logits = self.modules.ctc_lin(feats)
+        p_ctc = self.hparams.log_softmax(logits)
+        
+        if stage!= sb.Stage.TRAIN:
+            p_tokens = sb.decoders.ctc_greedy_decode(
+                p_ctc, wav_lens, blank_id=self.hparams.blank_index
+            )
+        else : 
+            p_tokens = None
+        return p_ctc, wav_lens, p_tokens
+    
+    
+    def treat_wav(self,sig):
+        multi_langual_feats = middle_layer(sig.to("cpu"), torch.tensor([1]).to("cpu"))
+        multi_langual_feats= multi_langual_feats.to(device)
+        feats, _ = self.modules.enc(multi_langual_feats)
+        logits = self.modules.ctc_lin(feats)
+        p_ctc = self.hparams.log_softmax(logits)
+        predicted_words =[]
+        for logs in p_ctc:
+            text = decoder.decode(logs.detach().cpu().numpy())
+            predicted_words.append(text.split(" "))
+        return " ".join(predicted_words[0])
+        
+
+    def compute_objectives(self, predictions, batch, stage):
+        """Computes the loss (CTC) given predictions and targets."""
+
+        p_ctc, wav_lens , predicted_tokens= predictions
+
+        ids = batch.id
+        tokens, tokens_lens = batch.tokens
+
+        loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
+
+
+        if stage == sb.Stage.VALID:
+            predicted_words = [
+                "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                for utt_seq in predicted_tokens
+            ]
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+        if stage ==sb.Stage.TEST : 
+            if self.hparams.language_modelling:
+                predicted_words = []
+                for logs in p_ctc:
+                    text = decoder.decode(logs.detach().cpu().numpy())
+                    predicted_words.append(text.split(" "))
+            else : 
+                predicted_words = [
+                    "".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
+                    for utt_seq in predicted_tokens
+                ]
+
+            target_words = [wrd.split(" ") for wrd in batch.wrd]
+            self.wer_metric.append(ids, predicted_words, target_words)
+            self.cer_metric.append(ids, predicted_words, target_words)
+
+        return loss
+
+    def fit_batch(self, batch):
+        """Train the parameters given a single batch in input"""
+        should_step = self.step % self.grad_accumulation_factor == 0
+        # Managing automatic mixed precision
+        # TOFIX: CTC fine-tuning currently is unstable
+        # This is certainly due to CTC being done in fp16 instead of fp32
+        if self.auto_mix_prec:
+            with torch.cuda.amp.autocast():
+                with self.no_sync():
+                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+            with self.no_sync(not should_step):
+                self.scaler.scale(
+                    loss / self.grad_accumulation_factor
+                ).backward()
+            if should_step:
+
+
+                self.scaler.unscale_(self.model_optimizer)
+                if self.check_gradients(loss):
+                    self.scaler.step(self.model_optimizer)
+                self.scaler.update()
+                self.zero_grad()
+                self.optimizer_step += 1
+        else:
+            # This is mandatory because HF models have a weird behavior with DDP
+            # on the forward pass
+            with self.no_sync():
+                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
+
+            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
+
+            with self.no_sync(not should_step):
+                (loss / self.grad_accumulation_factor).backward()
+            if should_step:
+                if self.check_gradients(loss):
+                    self.model_optimizer.step()
+                self.zero_grad()
+                self.optimizer_step += 1
+
+        self.on_fit_batch_end(batch, outputs, loss, should_step)
+        return loss.detach().cpu()
+
+    def evaluate_batch(self, batch, stage):
+        """Computations needed for validation/test batches"""
+        predictions = self.compute_forward(batch, stage=stage)
+        with torch.no_grad():
+            loss = self.compute_objectives(predictions, batch, stage=stage)
+        return loss.detach()
+
+    def on_stage_start(self, stage, epoch):
+        """Gets called at the beginning of each epoch"""
+        if stage != sb.Stage.TRAIN:
+            self.cer_metric = self.hparams.cer_computer()
+            self.wer_metric = self.hparams.error_rate_computer()
+
+    def on_stage_end(self, stage, stage_loss, epoch):
+        """Gets called at the end of an epoch."""
+        # Compute/store important stats
+        stage_stats = {"loss": stage_loss}
+        if stage == sb.Stage.TRAIN:
+            self.train_stats = stage_stats
+        else:
+            stage_stats["CER"] = self.cer_metric.summarize("error_rate")
+            stage_stats["WER"] = self.wer_metric.summarize("error_rate")
+
+        # Perform end-of-iteration things, like annealing, logging, etc.
+        if stage == sb.Stage.VALID:
+            old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
+                stage_stats["loss"]
+            )
+            sb.nnet.schedulers.update_learning_rate(
+                self.model_optimizer, new_lr_model
+            )
+            self.hparams.train_logger.log_stats(
+                stats_meta={
+                    "epoch": epoch,
+                    "lr_model": old_lr_model,
+                },
+                train_stats=self.train_stats,
+                valid_stats=stage_stats,
+            )
+            self.checkpointer.save_and_keep_only(
+                meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
+            )
+        elif stage == sb.Stage.TEST:
+            self.hparams.train_logger.log_stats(
+                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
+                test_stats=stage_stats,
+            )
+            with open(self.hparams.wer_file, "w") as w:
+                self.wer_metric.write_stats(w)
+
+    def init_optimizers(self):
+
+        self.model_optimizer = self.hparams.model_opt_class(
+            self.hparams.model.parameters()
+        )
+
+        if self.checkpointer is not None:
+            self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
+
+    def zero_grad(self, set_to_none=False):
+
+        self.model_optimizer.zero_grad(set_to_none)
+
+
+
+
+hparams_file, run_opts, overrides = sb.parse_arguments(["cs.yaml"])
+
+# If distributed_launch=True then
+# create ddp_group with the right communication protocol
+sb.utils.distributed.ddp_init_group(run_opts)
+
+with open(hparams_file) as fin:
+    hparams = load_hyperpyyaml(fin, overrides)
+
+# Create experiment directory
+sb.create_experiment_directory(
+    experiment_directory=hparams["output_folder"],
+    hyperparams_to_save=hparams_file,
+    overrides=overrides,
+)
+def read_labels_file(labels_file):
+    with open(labels_file, "r",encoding="utf-8") as lf:
+        lines = lf.read().splitlines()
+        division = "==="
+        numbers = {}
+        for line in lines :
+            if division in line :
+                break
+            string, number = line.split("=>")
+            number = int(number)
+            string = string[1:-2]
+            numbers[number] = string
+        return [numbers[x] for x in range(len(numbers))]
+
+label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
+special_labels = {
+    "blank_label": hparams["blank_index"],
+    "unk_label": hparams["unk_index"]
+}
+label_encoder.load_or_create(
+    path=lab_enc_file,
+    from_didatasets=[[]],
+    output_key="char_list",
+    special_labels=special_labels,
+    sequence_input=True,
+)
+
+
+labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
+labels = [""] + labels[1:-1] + ["1"]
+if hparams["language_modelling"]:
+    decoder = build_ctcdecoder(
+        labels,
+        kenlm_model_path=hparams["ngram_lm_path"],  # either .arpa or .bin file
+        alpha=0.5,  # tuned on a val set
+        beta=1,  # tuned on a val set
+    )
+
+
+
+run_opts["device"]="cpu"
+
+mixer = Mixer(
+    modules=hparams["modules"],
+    hparams=hparams,
+    run_opts=run_opts,
+    checkpointer=hparams["checkpointer"],
+)
+mixer.tokenizer = label_encoder
+mixer.device = "cpu"
+mixer.checkpointer.recover_if_possible()
+mixer.modules.eval()
+
+
+label_encoder = sb.dataio.encoder.CTCTextEncoder()
+
+
+# We dynamicaly add the tokenizer to our brain class.
+# NB: This tokenizer corresponds to the one used for the LM!!
+
+decoder = build_ctcdecoder(
+    labels,
+    kenlm_model_path= "arpas/everything.arpa",  # either .arpa or .bin file
+    alpha=0.5,  # tuned on a val set
+    beta=1,  # tuned on a val set
+)
+
+
+
+device = "cpu"
+mixer.device= "cpu"
+mixer.modules.to("cpu")
+
+from enum import Enum, auto
+class Stage(Enum):
+    TRAIN = auto()
+    VALID = auto()
+    TEST = auto()
+
+asr_brain.on_evaluate_start()
+asr_brain.modules.eval()
+
+
+import gradio as gr
+
+def treat_wav_file(file_mic,file_upload ,asr=mixer, device="cpu") :
+    if (file_mic is not None) and (file_upload is not None):
+        warn_output = "WARNING: You've uploaded an audio file and used the microphone. The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
+        wav = file_mic
+    elif (file_mic is None) and (file_upload is None):
+        return "ERROR: You have to either use the microphone or upload an audio file"
+    elif file_mic is not None:
+        wav = file_mic
+    else:
+        wav = file_upload
+    sig, sr = torchaudio.load(wav)
+    tensor_wav = sig.to(device)
+    resampled = torchaudio.functional.resample( tensor_wav, sr, 16000)
+    sentence = asr.treat_wav(resampled)
+    return sentence
+
+gr.Interface(
+    fn=treat_wav_file,
+    inputs=[gr.Audio(source="microphone", type='filepath', label = "record", optional = True),
+            gr.Audio(source="upload", type='filepath', label="filein", optional=True)]
+    ,outputs="text").launch()
+

results/non_semi_final_stac/env.log

@@ -0,0 +1,479 @@
+SpeechBrain system description
+==============================
+Python version:
+3.8.5 (default, Sep  4 2020, 07:30:14) 
+[GCC 7.3.0]
+==============================
+Installed Python packages:
+abkhazia==1.0
+absl-py==0.11.0
+aiofiles==23.2.1
+aiohttp==3.8.0
+aiosignal==1.2.0
+alabaster==0.7.12
+alembic==1.7.4
+altair==4.2.0
+altgraph==0.17
+antlr4-python3-runtime==4.9.3
+anyio==3.6.2
+appdirs==1.4.4
+argcomplete==1.12.2
+argon2-cffi==20.1.0
+arrow==1.2.3
+asgiref==3.6.0
+asteroid-filterbanks==0.4.0
+astunparse==1.6.3
+async-generator==1.10
+async-timeout==4.0.0
+attrdict==2.0.1
+attrs==20.3.0
+audeer==1.16.0
+audformat==0.11.5
+audinterface==0.7.0
+audiofile==1.0.0
+audiomentations==0.25.0
+audioread==2.1.9
+audobject==0.4.14
+audresample==0.1.6
+-e git+https://github.com/facebookresearch/WavAugment.git@54afcdb00ccc852c2f030f239f8532c9562b550e#egg=augment
+autopage==0.4.0
+Babel==2.9.0
+backcall==0.2.0
+backports.cached-property==1.0.2
+beautifulsoup4==4.10.0
+black==19.10b0
+bleach==3.3.0
+blessed==1.20.0
+boto3==1.20.2
+botocore==1.23.2
+bpemb==0.3.4
+braceexpand==0.1.7
+cachetools==4.2.0
+certifi @ file:///croot/certifi_1671487769961/work/certifi
+cffi==1.14.3
+cfgv==3.2.0
+chardet==3.0.4
+charset-normalizer==2.0.7
+click==7.1.2
+cliff==3.9.0
+clldutils==3.5.4
+cloudpickle==2.2.1
+cmaes==0.8.2
+cmake==3.18.4.post1
+cmd2==2.2.0
+colorama==0.4.4
+colorlog==4.6.2
+configparser==5.1.0
+conllu==4.5.3
+croniter==1.3.15
+cryptography==38.0.4
+csrgraph==0.1.28
+csvw==1.8.1
+cycler==0.10.0
+Cython==0.29.21
+dataclasses==0.6
+dateutils==0.6.12
+decorator==4.4.2
+deepdiff==6.3.0
+deepspeech==0.9.1
+defusedxml==0.7.1
+Deprecated==1.2.14
+dill==0.3.3
+Distance==0.1.3
+distlib==0.3.1
+Django==3.2.16
+django-auditlog==2.2.1
+django-filter==22.1
+django-js-asset==1.2.2
+django-mptt==0.14.0
+djangorestframework==3.14.0
+docker-pycreds==0.4.0
+docopt==0.6.2
+docutils==0.16
+drf-excel==2.2.0
+drf-flex-fields==1.0.0
+drf-renderer-xlsx==0.4.1
+easyocr==1.2.1
+editdistance==0.6.0
+einops==0.3.2
+emoji==2.2.0
+entrypoints==0.3
+et-xmlfile==1.1.0
+exceptiongroup==1.1.0
+farasapy==0.0.14
+fastapi==0.98.0
+fastjsonschema==2.17.1
+fasttext==0.9.2
+ffmpeg-python==0.2.0
+ffmpy==0.3.0
+filelock==3.0.12
+flair==0.12.2
+flake8==3.7.9
+flatbuffers==1.12
+frozendict==2.0.7
+frozenlist==1.2.0
+fsspec==2021.11.0
+ftfy==6.1.1
+future==0.18.2
+g2p-en==2.1.0
+gast==0.3.3
+gdown==4.4.0
+gdrive==0.1.5
+gensim==4.0.1
+gitdb==4.0.9
+GitPython==3.1.24
+google-api-core==2.11.1
+google-api-python-client==2.43.0
+google-auth==1.24.0
+google-auth-httplib2==0.1.0
+google-auth-oauthlib==0.5.3
+google-pasta==0.2.0
+googleapis-common-protos==1.59.1
+gradio==3.44.4
+gradio-client==0.5.1
+greenlet==1.1.2
+grpcio==1.32.0
+h11==0.14.0
+h5features==1.3.2
+h5py==2.10.0
+hierarchy==0.4.0
+hmmlearn==0.2.8
+htk-io==0.5
+httpcore==0.16.3
+httplib2==0.22.0
+httpx==0.23.3
+huggingface-hub==0.15.1
+hydra-colorlog==0.1.4
+hydra-core==1.3.2
+hyperopt==0.2.7
+HyperPyYAML==1.1.0
+hypothesis==6.61.2
+identify==1.5.10
+idna==2.10
+imageio==2.9.0
+imagesize==1.2.0
+importlib-metadata==4.8.1
+importlib-resources==5.2.2
+inflect==5.3.0
+inquirer==3.1.3
+ipadic==1.0.0
+ipyevents==2.0.1
+ipykernel==5.3.4
+ipython==7.19.0
+ipython-genutils==0.2.0
+ipywebrtc==0.6.0
+ipywidgets==7.6.3
+iso-639==0.4.5
+isodate==0.6.0
+isort==4.3.21
+itsdangerous==2.1.2
+Janome==0.5.0
+jedi==0.17.2
+jeepney==0.8.0
+jieba==0.42.1
+Jinja2==3.0.3
+jiwer==2.2.0
+jmespath==0.10.0
+joblib==0.17.0
+jsonschema==3.2.0
+julius==0.2.7
+jupyter-client==6.1.7
+jupyter-core==4.7.0
+jupyterlab-pygments==0.1.2
+jupyterlab-widgets==1.0.0
+kaitaistruct==0.9
+kaldi-io==0.9.4
+kaldi-python-io==1.2.2
+kaldiio==2.17.2
+kenlm @ https://github.com/kpu/kenlm/archive/master.zip
+Keras-Preprocessing==1.1.2
+kiwisolver==1.3.1
+lang-trans==0.6.0
+langdetect==1.0.9
+latexcodec==2.0.1
+ldap3==2.9.1
+librosa==0.9.0
+lightning-cloud==0.5.37
+lightning-utilities==0.8.0
+linkify-it-py==1.0.3
+lit==16.0.6
+llvmlite==0.35.0
+lxml==4.9.0
+Mako==1.1.5
+Markdown==3.3.3
+markdown-it-py==3.0.0
+MarkupSafe==2.1.3
+marshmallow==3.14.0
+matplotlib==3.3.3
+mccabe==0.6.1
+mcd==0.4
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+mecab-python3==1.0.3
+megatron-lm==2.2.0
+metrics==0.3.3
+mido==1.2.10
+mistune==0.8.4
+more-itertools==8.6.0
+mpld3==0.3
+mpmath==1.2.1
+multidict==5.2.0
+multiprocess==0.70.11.1
+nbclient==0.5.3
+nbconvert==5.6.1
+nbformat==5.9.0
+NEMO==4.3.2
+nemo-toolkit==1.4.0
+nest-asyncio==1.5.1
+networkx==2.8.8
+nltk==3.2.4
+nodeenv==1.5.0
+normalize==2.0.2
+notebook==6.3.0
+numba==0.52.0
+numpy==1.19.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-cupti-cu11==11.7.101
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+nvidia-cufft-cu11==10.9.0.58
+nvidia-curand-cu11==10.2.10.91
+nvidia-cusolver-cu11==11.4.0.1
+nvidia-cusparse-cu11==11.7.4.91
+nvidia-nccl-cu11==2.14.3
+nvidia-nvtx-cu11==11.7.91
+oauthlib==3.1.0
+omegaconf==2.3.0
+onnx==1.10.2
+OpenCC==1.1.2
+opencv-python==4.4.0.46
+openpyxl==3.0.9
+opensmile==2.2.0
+opt-einsum==3.3.0
+optuna==2.10.0
+ordered-set==4.1.0
+orjson==3.8.4
+oyaml==1.0
+packaging==22.0
+pandas==1.2.5
+pandocfilters==1.4.3
+pangu==4.0.6.1
+parameterized==0.8.1
+parso==0.7.1
+pathlib2==2.3.7.post1
+pathspec==0.5.5
+pathtools==0.1.2
+pbr==5.6.0
+pefile==2019.4.18
+pescador==2.1.0
+pesq==0.0.3
+pexpect==4.8.0
+phonemizer==2.2.1
+pickleshare==0.7.5
+Pillow==9.3.0
+pip-api==0.0.23
+pipreqs==0.4.11
+pluggy==0.13.1
+pooch==1.3.0
+portalocker==2.3.2
+pptree==3.1
+pre-commit==2.9.0
+preprocessing==0.1.13
+pretty-midi==0.2.9
+prettytable==2.2.1
+primePy==1.3
+progressbar2==3.53.1
+prometheus-client==0.10.1
+promise==2.3
+prompt-toolkit==3.0.8
+protobuf==3.20.3
+psutil==5.6.6
+ptyprocess==0.6.0
+py==1.9.0
+py-espeak-ng==0.1.8
+py4j==0.10.9.7
+pyannote.audio==2.1.1
+pyannote.core==4.5
+pyannote.database==4.1.3
+pyannote.metrics==3.2.1
+pyannote.pipeline==2.3
+pyannotebook==0.1.0.dev0
+PyArabic==0.6.15
+pyarrow==3.0.0
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pybind11==2.8.1
+pybtex==0.24.0
+pybtex-docutils==1.0.1
+pycodestyle==2.5.0
+pycparser==2.20
+pycryptodome==3.16.0
+pyctcdecode==0.4.0
+pydantic==1.10.4
+pyDeprecate==0.3.1
+pydub==0.25.1
+pyflakes==2.1.1
+Pygments==2.15.1
+pygtrie==2.5.0
+PyJWT==2.7.0
+pymodbus==2.5.3
+pyparsing==2.4.7
+pyperclip==1.8.2
+pypinyin==0.43.0
+pyrsistent==0.17.3
+pyserial==3.5
+PySocks==1.7.1
+pystoi==0.3.3
+pytest==5.4.1
+pytest-runner==5.3.1
+python-bidi==0.4.2
+python-crfsuite==0.9.7
+python-dateutil==2.8.2
+python-editor==1.0.4
+python-Levenshtein==0.12.2
+python-multipart==0.0.5
+python-utils==2.4.0
+pytorch-lightning==1.6.5
+pytorch-metric-learning==1.7.3
+pytorch-revgrad==0.2.0
+pytube==11.0.1
+pytz==2022.6
+PyWavelets==1.1.1
+PyYAML==6.0
+pyzmq==20.0.0
+rapidfuzz==1.8.2
+readchar==4.0.5
+regex==2020.11.13
+requests==2.28.1
+requests-oauthlib==1.3.0
+resampy==0.2.2
+rfc3986==1.4.0
+rich==13.4.2
+richenum==1.3.1
+rsa==4.7
+ruamel.yaml==0.17.21
+ruamel.yaml.clib==0.2.7
+s3m==1.1.0
+s3transfer==0.5.0
+sacrebleu==2.0.0
+sacremoses==0.0.44
+safetensors==0.3.1
+scikit-image==0.18.1
+scikit-learn==0.23.2
+scipy==1.5.4
+-e git+https://github.com/sanghack81/SDCIT@00d060dde733fde9345154a494f81e97fb395ca7#egg=SDCIT
+seaborn==0.11.1
+SecretStorage==3.3.3
+segments==2.1.3
+segtok==1.5.11
+semantic-version==2.10.0
+semver==2.13.0
+Send2Trash==1.5.0
+sentencepiece==0.1.99
+sentry-sdk==1.4.3
+shellingham==1.4.0
+shortuuid==1.0.7
+SIDEKIT==1.3.8.5.2
+simplejson==3.17.5
+singledispatchmethod==1.0
+six==1.15.0
+smart-open==5.0.0
+smmap==5.0.0
+sniffio==1.3.0
+snowballstemmer==2.0.0
+sortedcollections==2.1.0
+sortedcontainers==2.4.0
+sounddevice==0.4.5
+SoundFile==0.10.3.post1
+soupsieve==2.3
+sox==1.4.1
+sparsemax==0.1.9
+speechbrain==0.5.14
+sphfile==1.0.3
+Sphinx==3.3.1
+sphinx-rtd-theme==0.2.4
+sphinxcontrib-applehelp==1.0.2
+sphinxcontrib-bibtex==2.4.1
+sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-htmlhelp==1.0.3
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-serializinghtml==1.1.4
+SQLAlchemy==1.4.25
+sqlitedict==2.1.0
+sqlparse==0.4.2
+stanza==1.4.2
+starlette==0.27.0
+starsessions==1.3.0
+stevedore==3.4.0
+subprocess32==3.5.4
+sympy==1.9
+tabulate==0.8.9
+tensorboard==2.4.0
+tensorboard-plugin-wit==1.7.0
+tensorboardX==2.6.1
+tensorflow==2.4.0
+tensorflow-estimator==2.4.0
+termcolor==1.1.0
+terminado==0.9.4
+testpath==0.4.4
+threadpoolctl==2.1.0
+tifffile==2020.12.8
+tikzplotlib==0.9.8
+tinycss2==1.2.1
+tkseem==0.0.3
+tokenizers==0.13.3
+toml==0.10.2
+toolz==0.12.0
+torch==1.13.1
+torch-audiomentations==0.11.0
+torch-pitch-shift==1.2.4
+torch-stft==0.1.4
+torchaudio==0.13.1
+torchmetrics==0.11.4
+torchvision==0.14.1
+tornado==6.1
+tqdm==4.61.1
+trackrip==1.2.1
+traitlets==5.9.0
+transformer-smaller-training-vocab==0.3.1
+transformers==4.30.2
+triton==2.0.0
+typed-ast==1.4.1
+typer==0.4.0
+typing-extensions==4.4.0
+uc-micro-py==1.0.1
+Unidecode==1.3.2
+uritemplate==3.0.1
+urllib3==1.26.2
+uvicorn==0.20.0
+versioneer==0.28
+virtualenv==20.2.1
+wandb==0.12.6
+wcwidth==0.2.5
+webdataset==0.1.62
+webencodings==0.5.1
+websocket-client==1.6.1
+websockets==10.4
+Werkzeug==1.0.1
+wget==3.2
+widgetsnbextension==3.5.1
+Wikipedia-API==0.6.0
+wordninja==2.0.0
+wrapt==1.12.1
+xmltodict==0.13.0
+xxhash==2.0.0
+yamllint==1.23.0
+yarg==0.1.9
+yarl==1.7.2
+yaspin==2.1.0
+youtokentome==1.0.6
+youtube-dl==2021.6.6
+zipp==3.6.0
+==============================
+Git revision:
+be9098b
+==============================
+CUDA version:
+11.7

results/non_semi_final_stac/hyperparams.yaml

@@ -1,5 +1,5 @@
-# Generated 2023-09-20 from:
-# /home/salah/Code_Switched_Tunisian_Speech_Recognition/cs.yaml
+# Generated 2023-09-25 from:
+# /home/salah/Code-Switched-Tunisian-SpeechToText/cs.yaml
 # yamllint disable
 # Generated 2023-08-03 from:
 # /home/salah/new_tunisian_model/hparams/train_tunisian_withwavlm.yaml