wav2vec2-large-xlsr-marathi

Wav2Vec2-Large-XLSR-53-Marathi

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Marathi using the Open SLR64 dataset. When using this model, make sure that your speech input is sampled at 16kHz. This data contains only female voices but the model works well for male voices too. Trained on Google Colab Pro on Tesla P100 16GB GPU.<br> WER (Word Error Rate) on the Test Set: 12.70 %

Usage

The model can be used directly without a language model as follows, given that your dataset has Marathi actual_text and path_in_folder columns:

import torch, torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

#Since marathi is not present on Common Voice, script for reading the below dataset can be picked up from the eval script below
mr_test_dataset = all_data['test']

processor = Wav2Vec2Processor.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 
model = Wav2Vec2ForCTC.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 

resampler = torchaudio.transforms.Resample(48_000, 16_000) #first arg - input sample, second arg - output sample
# 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_in_folder"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch
mr_test_dataset = mr_test_dataset.map(speech_file_to_array_fn)
inputs = processor(mr_test_dataset["speech"][:5], 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:", mr_test_dataset["actual_text"][:5])

Evaluation

Evaluated on 10% of the Marathi data on Open SLR-64.

import os, re, torch, torchaudio
from datasets import Dataset, load_metric
import pandas as pd
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

#below is a custom script to be used for reading marathi dataset since its not present on the Common Voice
dataset_path = "./OpenSLR-64_Marathi/mr_in_female/" #TODO : include the path of the dataset extracted from http://openslr.org/64/
audio_df = pd.read_csv(os.path.join(dataset_path,'line_index.tsv'),sep='\t',header=None)
audio_df.columns = ['path_in_folder','actual_text']
audio_df['path_in_folder'] = audio_df['path_in_folder'].apply(lambda x: dataset_path + x + '.wav')
audio_df = audio_df.sample(frac=1, random_state=2020).reset_index(drop=True) #seed number is important for reproducibility of WER score
all_data = Dataset.from_pandas(audio_df)
all_data = all_data.train_test_split(test_size=0.10,seed=2020) #seed number is important for reproducibility of WER score

mr_test_dataset = all_data['test']
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi")
model = Wav2Vec2ForCTC.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“]' 
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["actual_text"] = re.sub(chars_to_ignore_regex, '', batch["actual_text"]).lower()
  speech_array, sampling_rate = torchaudio.load(batch["path_in_folder"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch
mr_test_dataset = mr_test_dataset.map(speech_file_to_array_fn)
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 = mr_test_dataset.map(evaluate, batched=True, batch_size=8)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["actual_text"])))

Training

Train-Test ratio was 90:10. The training notebook Colab link here.

Training Config and Summary

weights-and-biases run summary here