xlsr-bengali

README.md

@@ -0,0 +1,99 @@
+---
+language: Bengali
+datasets:
+- OpenSLR
+metrics:
+- wer
+tags:
+- audio
+- automatic-speech-recognition
+- speech
+- xlsr-fine-tuning-week
+license: Attribution-ShareAlike 4.0 International
+model-index:
+- name: XLSR Wav2Vec2 Bengali by Tanmoy Sarkar
+  results:
+  - task: 
+      name: Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: OpenSLR
+      type: OpenSLR
+      args: or
+    metrics:
+       - name: Test WER
+         type: wer
+         value: 88.58
+---
+# Wav2Vec2-Large-XLSR-Bengali
+Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) Bengali using the [Bengali ASR training data set containing ~196K utterances](https://www.openslr.org/53/).
+When using this model, make sure that your speech input is sampled at 16kHz.
+## Usage
+Dataset must be downloaded from [this website](https://www.openslr.org/53/) and preprocessed accordingly. For example 1250 test samples has been chosen.
+```python
+import pandas as pd
+test_dataset = pd.read_csv('utt_spk_text.tsv', sep='\t', header=None)[60000:61250]
+test_dataset.columns = ["audio_path", "__", "label"]
+test_dataset = test_data.drop("__", axis=1)
+def add_file_path(text):
+  path = "data/" + text[:2] + "/" + text + '.flac'
+  return path
+test_dataset['audio_path'] = test_dataset['audio_path'].map(lambda x: add_file_path(x))
+```
+The model can be used directly (without a language model) as follows:
+```python
+import torch
+import torchaudio
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+processor = Wav2Vec2Processor.from_pretrained("tanmoyio/wav2vec2-large-xlsr-bengali")
+model = Wav2Vec2ForCTC.from_pretrained("tanmoyio/wav2vec2-large-xlsr-bengali")
+resampler = torchaudio.transforms.Resample(48_000, 16_000)
+# Preprocessing the datasets.
+# We need to read the aduio files as arrays
+def speech_file_to_array_fn(batch):
+  speech_array, sampling_rate = torchaudio.load(batch["audio_path"])
+  batch["speech"] = resampler(speech_array).squeeze().numpy()
+  return batch
+test_dataset = test_dataset.map(speech_file_to_array_fn)
+inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
+with torch.no_grad():
+  logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
+predicted_ids = torch.argmax(logits, dim=-1)
+print("Prediction:", processor.batch_decode(predicted_ids))
+print("Reference:", test_dataset["label"][:2])
+```
+## Evaluation
+The model can be evaluated as follows on the Bengali test data of OpenSLR.  
+```python
+import torch
+import torchaudio
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+import re
+wer = load_metric("wer")
+processor = Wav2Vec2Processor.from_pretrained("tanmoyio/wav2vec2-large-xlsr-bengali")
+model = Wav2Vec2ForCTC.from_pretrained("tanmoyio/wav2vec2-large-xlsr-bengali")
+model.to("cuda")
+resampler = torchaudio.transforms.Resample(48_000, 16_000)
+# Preprocessing the datasets.
+# We need to read the aduio files as arrays
+def speech_file_to_array_fn(batch):
+  batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["label"]).lower()
+  speech_array, sampling_rate = torchaudio.load(batch["path"])
+  batch["speech"] = resampler(speech_array).squeeze().numpy()
+  return batch
+test_dataset = test_dataset.map(speech_file_to_array_fn)
+# Preprocessing the datasets.
+# We need to read the aduio files as arrays
+def evaluate(batch):
+  inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
+  with torch.no_grad():
+      logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
+      pred_ids = torch.argmax(logits, dim=-1)
+      batch["pred_strings"] = processor.batch_decode(pred_ids)
+      return batch
+result = test_dataset.map(evaluate, batched=True, batch_size=8)
+print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
+```
+**Test Result**: 88.58 %
+## Training
+The script used for training can be found [Bengali ASR Fine Tuning Wav2Vec2](https://colab.research.google.com/drive/1Bkc5C_cJV9BeS0FD0MuHyayl8hqcbdRZ?usp=sharing)

config.json

@@ -0,0 +1,76 @@
+{
+  "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
+  "activation_dropout": 0.0,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": true,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.0,
+  "final_dropout": 0.0,
+  "gradient_checkpointing": true,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.1,
+  "mask_channel_length": 10,
+  "mask_channel_min_space": 1,
+  "mask_channel_other": 0.0,
+  "mask_channel_prob": 0.0,
+  "mask_channel_selection": "static",
+  "mask_feature_length": 10,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_space": 1,
+  "mask_time_other": 0.0,
+  "mask_time_prob": 0.05,
+  "mask_time_selection": "static",
+  "model_type": "wav2vec2",
+  "num_attention_heads": 16,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 24,
+  "pad_token_id": 118,
+  "transformers_version": "4.5.0.dev0",
+  "vocab_size": 119
+}

preprocessor_config.json

@@ -0,0 +1,8 @@
+{
+  "do_normalize": true,
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:788e64f971b140bbc265df0486041a0607fe55baf5a22c15b31140b5e5cc1813
+size 1262421719

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|"}

vocab.json

@@ -0,0 +1 @@
+{"র": 0, "0": 1, "t": 2, "০": 3, "‌": 4, "ন": 5, "প": 6, "৬": 7, "ঃ": 8, "\t": 9, "ড": 10, "হ": 11, "ব": 12, "8": 13, "১": 14, "1": 15, "৮": 16, "স": 17, "ো": 18, "ঙ": 19, "৭": 20, "o": 21, "ূ": 22, "ৈ": 23, "‍": 24, "d": 25, "ফ": 26, "য়": 27, "c": 28, "ঊ": 29, "৪": 30, "b": 31, "জ": 32, "p": 33, "9": 34, "3": 35, "ড়": 36, "2": 37, "w": 39, "ল": 40, "ু": 41, "ঝ": 42, "\"": 43, "ই": 44, "ঘ": 45, "n": 46, "ক": 47, "ট": 48, "r": 49, "ণ": 50, "ছ": 51, "থ": 52, "ভ": 53, "শ": 54, ":": 55, "দ": 56, "6": 57, "a": 58, "ম": 59, "-": 60, "এ": 61, "!": 62, "ত": 63, "খ": 64, "চ": 65, ",": 66, ".": 67, "২": 68, "l": 69, "আ": 70, "গ": 71, "্": 72, "4": 73, "ঞ": 74, "৩": 75, "7": 76, "য": 77, "s": 78, "/": 79, "”": 80, "়": 81, "ও": 82, "e": 83, "ৌ": 84, "ঈ": 85, "'": 86, "ে": 87, "া": 88, "ঢ": 89, "\n": 90, "ঐ": 91, "ঔ": 92, "‘": 93, "উ": 94, "’": 95, "ধ": 96, "g": 97, "ঠ": 98, "ং": 99, "।": 100, "u": 101, "5": 102, "ৎ": 103, "v": 104, "ঋ": 105, "ী": 106, "৯": 107, "ঁ": 108, "অ": 109, "f": 110, "ষ": 111, "“": 112, "৫": 113, "ৃ": 114, "?": 115, "ি": 116, "|": 38, "[UNK]": 117, "[PAD]": 118}