Updated model

.ipynb_checkpoints/README-checkpoint.md

@@ -29,9 +29,12 @@ results:
         type: iiith
         args: hi
     metrics:
-       - name: Test WER
+       - name: Custom Dataset Hindi WER
          type: wer
-         value: 19.05
+         value: 17.23
+       - name: CommonVoice Hindi (Test) WER 
+         type: wer
+         value: 56.46
 ---
 
 # Wav2Vec2-Large-XLSR-53-Hindi
@@ -41,11 +44,11 @@ Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav
 - [Indic TTS- IITM](https://www.iitm.ac.in/donlab/tts/index.php) and
 - [IIITH - Indic Speech Datasets](http://speech.iiit.ac.in/index.php/research-svl/69.html)
 
-The Hindi CommonVoice data is skewed towards male voices. However the other Indic datasets are well balanced. 
+The Indic datasets are well balanced across gender and accents. However the CommonVoice dataset is skewed towards male voices
 
-Fine-tuned on facebook/wav2vec2-large-xlsr-53 using Hindi dataset :: 30 epochs >> 19.05% WER
-                          Resuming from checkpoints trained for another XX epochs >> XX.XX%
 
+Fine-tuned on facebook/wav2vec2-large-xlsr-53 using Hindi dataset :: 60 epochs >> 17.05% WER
+                          
 When using this model, make sure that your speech input is sampled at 16kHz.
 
 ## Usage
@@ -85,7 +88,81 @@ print("Reference:", test_dataset["sentence"][:2])
 
 ## Evaluation
 
-The model can be evaluated as follows on the Hindi test data of Common Voice. 
+The model can be evaluated as follows on the following two datasets: 
+1. Custom dataset created from 20% of Indic, IIITH and CV (test): 17. 
+2. CommonVoice Hindi test dataset
+
+```python
+import torch
+import torchaudio
+from datasets import load_dataset, load_metric
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+import re
+
+## Load the datasets
+test_dataset = load_dataset("common_voice", "hi", split="test") 
+
+indic = load_dataset("csv", data_files= {'train':"/workspace/data/hi2/indic_train_full.csv",
+                                        "test": "/workspace/data/hi2/indic_test_full.csv"}, download_mode="force_redownload")
+iiith = load_dataset("csv", data_files= {"train": "/workspace/data/hi2/iiit_hi_train.csv", 
+                                        "test": "/workspace/data/hi2/iiit_hi_test.csv"}, download_mode="force_redownload")
+
+## Pre-process datasets and concatenate to create test dataset
+# Drop columns of common_voice
+split = ['train', 'test', 'validation', 'other', 'invalidated']
+
+for sp in split:
+    common_voice[sp] = common_voice[sp].remove_columns(['client_id', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment']) 
+    
+common_voice = common_voice.rename_column('path', 'audio_path')
+common_voice = common_voice.rename_column('sentence', 'target_text')
+
+train_dataset = datasets.concatenate_datasets([indic['train'], iiith['train'], common_voice['train']])
+test_dataset = datasets.concatenate_datasets([indic['test'], iiith['test'], common_voice['test'], common_voice['validation']])
+
+## Load model from HF hub
+
+wer = load_metric("wer")
+
+processor = Wav2Vec2Processor.from_pretrained("skylord/wav2vec2-large-xlsr-hindi") 
+model = Wav2Vec2ForCTC.from_pretrained("skylord/wav2vec2-large-xlsr-hindi")
+model.to("cuda")
+
+chars_to_ignore_regex = '[\,\?\.\!\-\'\;\:\"\“\%\‘\”\�Utrnle\_]'
+unicode_ignore_regex = r'[dceMaWpmFui\xa0\u200d]' # Some unwanted unicode chars
+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["target_text"] = re.sub(chars_to_ignore_regex, '', batch["target_text"])
+  batch["target_text"] = re.sub(unicode_ignore_regex, '', batch["target_text"])
+    
+  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)
+
+# 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 on custom dataset**: 17.23 % 
+
+
 
 
 ```python
@@ -132,7 +209,8 @@ 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**: 19.056 % 
+**Test Result on CommonVoice**: 56.46 % 
+
 
 
 ## Training

README.md

@@ -34,7 +34,7 @@ results:
          value: 17.23
        - name: CommonVoice Hindi (Test) WER 
          type: wer
-         value: 52.35
+         value: 56.46
 ---
 
 # Wav2Vec2-Large-XLSR-53-Hindi
@@ -44,11 +44,11 @@ Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav
 - [Indic TTS- IITM](https://www.iitm.ac.in/donlab/tts/index.php) and
 - [IIITH - Indic Speech Datasets](http://speech.iiit.ac.in/index.php/research-svl/69.html)
 
-The Hindi CommonVoice data is skewed towards male voices. However the other Indic datasets are well balanced. 
+The Indic datasets are well balanced across gender and accents. However the CommonVoice dataset is skewed towards male voices
 
-Fine-tuned on facebook/wav2vec2-large-xlsr-53 using Hindi dataset :: 30 epochs >> 19.05% WER
-                          Resuming from checkpoints trained for another XX epochs >> XX.XX%
 
+Fine-tuned on facebook/wav2vec2-large-xlsr-53 using Hindi dataset :: 60 epochs >> 17.05% WER
+                          
 When using this model, make sure that your speech input is sampled at 16kHz.
 
 ## Usage
@@ -89,7 +89,7 @@ print("Reference:", test_dataset["sentence"][:2])
 ## Evaluation
 
 The model can be evaluated as follows on the following two datasets: 
-1. Custom dataset created from 20% of Indic, IIITH and CV (test)
+1. Custom dataset created from 20% of Indic, IIITH and CV (test): 17. 
 2. CommonVoice Hindi test dataset
 
 ```python
@@ -160,7 +160,7 @@ 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 on custom dataset**: 19.xx % 
+**Test Result on custom dataset**: 17.23 % 
 
 
 
@@ -209,7 +209,7 @@ 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 on CommonVoice**: 52.xx % 
+**Test Result on CommonVoice**: 56.46 % 
 
 
 

config.json

@@ -70,7 +70,7 @@
   "num_conv_pos_embeddings": 128,
   "num_feat_extract_layers": 7,
   "num_hidden_layers": 24,
-  "pad_token_id": 93,
+  "pad_token_id": 74,
   "transformers_version": "4.5.0.dev0",
-  "vocab_size": 94
+  "vocab_size": 75
 }

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:be0fd803682bd53a5e10531a8c98d1d154a4303d6d03d28a07c09b221c0e324e
-size 1262319255
+oid sha256:c42f7b3bd5cb4694035320ae412c100449b231a49a05e2194f8cc844646fc697
+size 1262241303

vocab.json

@@ -1 +1 @@
-{"प": 0, "ण": 1, "ड": 2, "फ": 3, "फ़": 4, "ं": 5, "?": 6, "ष": 7, "ँ": 8, "ृ": 9, "थ": 10, "ऊ": 11, "u": 12, "ग": 13, "r": 14, "द": 15, "र": 16, "ी": 17, "झ": 18, "ः": 19, "ब": 20, "ऐ": 21, ".": 22, "ख": 23, "घ": 24, "स": 25, "ौ": 26, "ई": 27, "F": 28, "उ": 29, "छ": 30, "ऋ": 31, "ि": 32, "े": 33, "!": 34, "ू": 35, "न": 37, "आ": 38, "ै": 39, "ऑ": 40, "भ": 41, ",": 42, "्": 43, "य": 44, "ञ": 45, "ढ": 46, "इ": 47, "।": 48, "क": 49, "ऱ": 50, "ङ": 51, "ॅ": 52, "ओ": 53, "ज़": 54, "ॉ": 55, "e": 56, "p": 57, "a": 58, "l": 59, "M": 60, "ध": 61, "ा": 62, "व": 63, "ट": 64, ":": 65, "\"": 66, "'": 67, "ढ़": 68, "ड़": 69, "ो": 70, " ": 71, "ल": 72, "अ": 73, "क़": 74, "त": 75, "ु": 76, "औ": 77, "m": 78, "श": 79, "च": 80, "़": 81, "ए": 82, "ह": 83, "W": 84, "ग़": 85, "ठ": 86, "ज": 87, "-": 88, "i": 89, "ख़": 90, "म": 91, "/": 36, "[UNK]": 92, "[PAD]": 93}
+{"प": 0, "ग": 1, "ौ": 2, "घ": 3, "झ": 4, "ख़": 5, "ष": 6, "छ": 7, "च": 8, "ढ़": 9, "ॅ": 10, "ऱ": 11, "ा": 12, "ख": 13, "ँ": 14, "ड": 15, "ह": 16, "श": 17, "ि": 18, "उ": 19, "ो": 20, "ऊ": 21, "ॉ": 22, "ग़": 23, "ऋ": 24, "्": 25, "ः": 26, "ं": 27, "द": 28, "ळ": 29, "ठ": 30, "ृ": 31, "ल": 32, "ज़": 33, "क": 34, "त": 35, "ध": 36, "ई": 37, "र": 38, "म": 39, "ज": 41, "आ": 42, "भ": 43, "।": 44, "इ": 45, "फ़": 46, "ञ": 47, "य": 48, "ऐ": 49, "ऑ": 50, "ट": 51, "ङ": 52, "थ": 53, "ी": 54, "ु": 55, "े": 56, "ै": 57, "न": 58, "स": 59, "ू": 60, "ब": 61, "ओ": 62, "व": 63, "ड़": 64, "अ": 65, "औ": 66, "़": 67, "क़": 68, "फ": 69, "ढ": 70, "ए": 71, "ण": 72, "/": 40, "[UNK]": 73, "[PAD]": 74}