skylord/wav2vec2-large-xlsr-hindi

Wav2Vec2-Large-XLSR-53-Hindi

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Hindi using the following datasets:

• Common Voice,
• Indic TTS- IITM and
• IIITH - Indic Speech Datasets

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 :: 60 epochs >> 17.05% WER

When using this model, make sure that your speech input is sampled at 16kHz.

Usage

The model can be used directly (without a language model) as follows:

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
test_dataset = load_dataset("common_voice", "hi", split="test")

processor = Wav2Vec2Processor.from_pretrained("skylord/wav2vec2-large-xlsr-hindi") 
model = Wav2Vec2ForCTC.from_pretrained("skylord/wav2vec2-large-xlsr-hindi") 

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["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["sentence"][:2])

Predictions

*Some good ones ..... *

Predictions	Reference
फिर वो सूरज तारे पहाड बारिश पदछड़ दिन रात शाम नदी बर्फ़ समुद्र धुंध हवा कुछ भी हो सकती है	फिर वो सूरज तारे पहाड़ बारिश पतझड़ दिन रात शाम नदी बर्फ़ समुद्र धुंध हवा कुछ भी हो सकती है
इस कारण जंगल में बडी दूर स्थित राघव के आश्रम में लोघ कम आने लगे और अधिकांश भक्त सुंदर के आश्रम में जाने लगे	इस कारण जंगल में बड़ी दूर स्थित राघव के आश्रम में लोग कम आने लगे और अधिकांश भक्त सुन्दर के आश्रम में जाने लगे
अपने बचन के अनुसार शुभमूर्त पर अनंत दक्षिणी पर्वत गया और मंत्रों का जप करके सरोवर में उतरा	अपने बचन के अनुसार शुभमुहूर्त पर अनंत दक्षिणी पर्वत गया और मंत्रों का जप करके सरोवर में उतरा

*Some crappy stuff .... *

Predictions	Reference
वस गनिल साफ़ है।	उसका दिल साफ़ है।
चाय वा एक कुछ लैंगे हब	चायवाय कुछ लेंगे आप
टॉम आधे है स्कूल हें है	टॉम अभी भी स्कूल में है

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): WER 17.xx%
2. CommonVoice Hindi test dataset: WER 56.xx%

Links to the datasets are provided above (check the links at the start of the README)

train-test csv files are shared on the following gdrive links: a. IIITH train test b. Indic TTS train test

Update the audio_path as per your local file structure.

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 %

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

test_dataset = load_dataset("common_voice", "hi", split="test") 
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["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).sub(unicode_ignore_regex, '', batch["sentence"])
  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 on CommonVoice: 56.46 %

Training

The Common Voice train, validation, datasets were used for training as well as

The script used for training & wandb dashboard can be found here