Upload 8 files

README.md

@@ -0,0 +1,25 @@
+---
+tags:
+- automatic-speech-recognition
+library_name: generic
+---
+
+# Automatic Speech Recognition repository template
+
+This is a template repository for Automatic Speech Recognition to support generic inference with Hugging Face Hub generic Inference API. There are two required steps:
+
+1. Specify the requirements by defining a `requirements.txt` file.
+2. Implement the `pipeline.py` `__init__` and `__call__` methods. These methods are called by the Inference API. The `__init__` method should load the model and preload all the elements needed for inference (model, processors, tokenizers, etc.). This is only called once. The `__call__` method performs the actual inference. Make sure to follow the same input/output specifications defined in the template for the pipeline to work.
+
+Example repos
+* https://huggingface.co/osanseviero/pyctcdecode_asr
+
+## How to start
+First create a repo in https://hf.co/new. 
+Then clone this template and push it to your repo.
+```
+git clone https://huggingface.co/templates/automatic-speech-recognition
+cd automatic-speech-recognition
+git remote set-url origin https://huggingface.co/$YOUR_USER/$YOUR_REPO_NAME
+git push --force
+```

config.json

@@ -0,0 +1,79 @@
+{
+  "_name_or_path": "facebook/wav2vec2-large-robust",
+  "activation_dropout": 0.1,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 768,
+  "contrastive_logits_temperature": 0.1,
+  "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": false,
+  "diversity_loss_weight": 0.1,
+  "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": 0.1,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.1,
+  "mask_feature_length": 10,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_prob": 0.05,
+  "model_type": "wav2vec2",
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 24,
+  "num_negatives": 100,
+  "pad_token_id": 0,
+  "proj_codevector_dim": 768,
+  "torch_dtype": "float32",
+  "transformers_version": "4.12.5",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 71
+}

pipeline.py

@@ -0,0 +1,88 @@
+from typing import  Dict
+import numpy as np
+from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
+import torch
+
+
+#define groups
+#make sure that all phonemes covered in each group
+g1 = ['p_alveolar','n_alveolar']
+g2 = ['p_palatal','n_palatal']
+g3 = ['p_dental','n_dental']
+g4 = ['p_glottal','n_glottal']
+g5 = ['p_labial','n_labial']
+g6 = ['p_velar','n_velar']
+g7 = ['p_anterior','n_anterior']
+g8 = ['p_posterior','n_posterior']
+g9 = ['p_retroflex','n_retroflex']
+g10 = ['p_mid','n_mid']
+g11 = ['p_high_v','n_high_v']
+g12 = ['p_low','n_low']
+g13 = ['p_front','n_front']
+g14 = ['p_back','n_back']
+g15 = ['p_central','n_central']
+g16 = ['p_consonant','n_consonant']
+g17 = ['p_sonorant','n_sonorant']
+g18 = ['p_long','n_long']
+g19 = ['p_short','n_short']
+g20 = ['p_vowel','n_vowel']
+g21 = ['p_semivowel','n_semivowel']
+g22 = ['p_fricative','n_fricative']
+g23 = ['p_nasal','n_nasal']
+g24 = ['p_stop','n_stop']
+g25 = ['p_approximant','n_approximant']
+g26 = ['p_affricate','n_affricate']
+g27 = ['p_liquid','n_liquid']
+g28 = ['p_continuant','n_continuant']
+g29 = ['p_monophthong','n_monophthong']
+g30 = ['p_diphthong','n_diphthong']
+g31 = ['p_round','n_round']
+g32 = ['p_voiced','n_voiced']
+g33 = ['p_bilabial','n_bilabial']
+g34 = ['p_coronal','n_coronal']
+g35 = ['p_dorsal','n_dorsal']
+groups = [g1,g2,g3,g4,g5,g6,g7,g8,g9,g10,g11,g12,g13,g14,g15,g16,g17,g18,g19,g20,g21,g22,g23,g24,g25,g26,g27,g28,g29,g30,g31,g32,g33,g34,g35]
+
+class PreTrainedPipeline():
+    def __init__(self, path=""):
+        # IMPLEMENT_THIS
+        # Preload all the elements you are going to need at inference.
+        # For instance your model, processors, tokenizer that might be needed.
+        # This function is only called once, so do all the heavy processing I/O here"""
+        self.sampling_rate = 16000
+        
+        self.processor = Wav2Vec2Processor.from_pretrained('.')
+        self.model = Wav2Vec2ForCTC.from_pretrained('.')
+        self.group_ids = [sorted(self.processor.tokenizer.convert_tokens_to_ids(group)) for group in groups]
+        self.group_ids = [dict([(x[0]+1,x[1]) for x in list(enumerate(g))]) for g in self.group_ids] #This is the inversion of the one used in training as here we need to map prediction back to original tokens
+        
+
+    def __call__(self, inputs: np.array)-> Dict[str, str]:
+        """
+        Args:
+            inputs (:obj:`np.array`):
+                The raw waveform of audio received. By default at 16KHz.
+        Return:
+            A :obj:`dict`:. The object return should be liked {"text": "XXX"} containing
+            the detected text from the input audio.
+        """
+        # IMPLEMENT_THIS
+        input_values = self.processor(audio=inputs, sampling_rate=self.sampling_rate, return_tensors="pt").input_values
+        
+        if torch.cuda.is_available():
+            self.model.to("cuda")
+            input_values = input_values.to("cuda")
+       
+        with torch.no_grad():
+            logits = self.model(input_values).logits
+            
+        mask = torch.zeros(logits.size()[2], dtype = torch.bool)
+        mask[0] = True
+        mask[list(self.group_ids[31].values())] = True
+        logits_g = logits[:,:,mask]
+        pred_ids = torch.argmax(logits_g,dim=-1)
+        pred_ids = pred_ids.cpu().apply_(lambda x: self.group_ids[31].get(x,x))
+        pred = self.processor.batch_decode(pred_ids,spaces_between_special_tokens=True)[0]
+        pred = pred.replace('p_','+').replace('n_', '-')
+        return({"text":pred})
+

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": false,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0275f1c9c019e5543cf3a95143aa762f1500b75b661bdf970b4ec7e221d363c
+size 1262214769

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": "", "tokenizer_class": "Wav2Vec2CTCTokenizer"}

vocab.json

@@ -0,0 +1 @@
+{"<pad>": 0, "n_affricate": 1, "n_alveolar": 2, "n_anterior": 3, "n_approximant": 4, "n_back": 5, "n_bilabial": 6, "n_central": 7, "n_consonant": 8, "n_continuant": 9, "n_coronal": 10, "n_dental": 11, "n_diphthong": 12, "n_dorsal": 13, "n_fricative": 14, "n_front": 15, "n_glottal": 16, "n_high_v": 17, "n_labial": 18, "n_liquid": 19, "n_long": 20, "n_low": 21, "n_mid": 22, "n_monophthong": 23, "n_nasal": 24, "n_palatal": 25, "n_posterior": 26, "n_retroflex": 27, "n_round": 28, "n_semivowel": 29, "n_short": 30, "n_sonorant": 31, "n_stop": 32, "n_velar": 33, "n_voiced": 34, "n_vowel": 35, "p_affricate": 36, "p_alveolar": 37, "p_anterior": 38, "p_approximant": 39, "p_back": 40, "p_bilabial": 41, "p_central": 42, "p_consonant": 43, "p_continuant": 44, "p_coronal": 45, "p_dental": 46, "p_diphthong": 47, "p_dorsal": 48, "p_fricative": 49, "p_front": 50, "p_glottal": 51, "p_high_v": 52, "p_labial": 53, "p_liquid": 54, "p_long": 55, "p_low": 56, "p_mid": 57, "p_monophthong": 58, "p_nasal": 59, "p_palatal": 60, "p_posterior": 61, "p_retroflex": 62, "p_round": 63, "p_semivowel": 64, "p_short": 65, "p_sonorant": 66, "p_stop": 67, "p_velar": 68, "p_voiced": 69, "p_vowel": 70}