metadata

language: mr
datasets:
  - openslr
  - interspeech_2021_asr
metrics:
  - wer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
  - hindi
  - marathi
license: apache-2.0
model-index:
  - name: XLSR Wav2Vec2 Large 53 Hindi-Marathi by Tanmay Laud
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: OpenSLR hi, OpenSLR mr
          type: openslr, interspeech_2021_asr
        metrics:
          - name: Test WER
            type: wer
            value: 24.92

Wav2Vec2-Large-XLSR-53-Hindi-Marathi

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Hindi and Marathi using the OpenSLR SLR64 datasets. 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, assuming you have a dataset with Marathi text and audio_path fields:

import torch
import torchaudio
import librosa
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# test_data = #TODO: WRITE YOUR CODE TO LOAD THE TEST DATASET. For sample see the Colab link in Training Section.

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

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["audio_path"])
    batch["speech"] = librosa.resample(speech_array[0].numpy(), sampling_rate, 16_000) # sampling_rate can vary
    return batch

test_data= test_data.map(speech_file_to_array_fn)
inputs = processor(test_data["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_data["text"][:2])
Evaluation
The model can be evaluated as follows on 10% of the Marathi data on OpenSLR.

import torchaudio
from datasets import load_metric
from transformers import Wav2Vec2Processor,Wav2Vec2ForCTC
import torch
import librosa
import numpy as np
import re

wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("tanmaylaud/wav2vec2-large-xlsr-hindi-marathi")
model = Wav2Vec2ForCTC.from_pretrained("tanmaylaud/wav2vec2-large-xlsr-hindi-marathi")

model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\।]'

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"])
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = speech_array[0].numpy()
    batch["sampling_rate"] = sampling_rate
    batch["target_text"] = batch["sentence"]
    batch["speech"] = librosa.resample(np.asarray(batch["speech"]), sampling_rate, 16_000)
    batch["sampling_rate"] = 16_000
    return batch

test= test.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the audio 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, group_tokens=False)
        # we do not want to group tokens when computing the metrics
        return batch

result = test.map(evaluate, batched=True, batch_size=32)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["text"])))

Link to eval notebook : https://colab.research.google.com/drive/1nZRTgKfxCD9cvy90wikTHkg2il3zgcqW#scrollTo=cXWFbhb0d7DT