Whisper Hindi Small

This model is a fine-tuned version of openai/whisper-small on the Hindi data available from multiple publicly available ASR corpuses. It has been fine-tuned as a part of the Whisper fine-tuning sprint.

NOTE: The code used to train this model is available for re-use in the whisper-finetune repository.

Usage

In order to evaluate this model on an entire dataset, the evaluation codes available in the whisper-finetune repository can be used.

The same repository also provides the scripts for faster inference using whisper-jax.

In order to infer a single audio file using this model, the following code snippet can be used:

>>> import torch
>>> from transformers import pipeline

>>> # path to the audio file to be transcribed
>>> audio = "/path/to/audio.format"
>>> device = "cuda:0" if torch.cuda.is_available() else "cpu"

>>> transcribe = pipeline(task="automatic-speech-recognition", model="vasista22/whisper-hindi-small", chunk_length_s=30, device=device)
>>> transcribe.model.config.forced_decoder_ids = transcribe.tokenizer.get_decoder_prompt_ids(language="hi", task="transcribe")

>>> print('Transcription: ', transcribe(audio)["text"])

For faster inference of whisper models, the whisper-jax library can be used. Please follow the necessary installation steps as mentioned here , before using the following code snippet:

>>> import jax.numpy as jnp
>>> from whisper_jax import FlaxWhisperForConditionalGeneration, FlaxWhisperPipline

>>> # path to the audio file to be transcribed
>>> audio = "/path/to/audio.format"

>>> transcribe = FlaxWhisperPipline("vasista22/whisper-hindi-small", batch_size=16)
>>> transcribe.model.config.forced_decoder_ids = transcribe.tokenizer.get_decoder_prompt_ids(language="hi", task="transcribe")

>>> print('Transcription: ', transcribe(audio)["text"])

Training and evaluation data

Training Data:

• GramVaani ASR Corpus
• ULCA ASR Corpus
• Shrutilipi ASR Corpus
• Google/Fleurs Train+Dev set

Evaluation Data:

• GramVaani ASR Corpus Test Set
• Google/Fleurs Test Set

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1.75e-05
• train_batch_size: 48
• eval_batch_size: 32
• seed: 22
• optimizer: adamw_bnb_8bit
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 20000
• training_steps: 19377 (Initially set to 129180 steps)
• mixed_precision_training: True

Acknowledgement

This work was done at Speech Lab, IIT Madras .

The compute resources for this work were funded by "Bhashini: National Language translation Mission" project of the Ministry of Electronics and Information Technology (MeitY), Government of India.