Whisper Medium Malayalam

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

WER: 38.6207
CER: 7.3256

Note that Whisper's normalization has major issues for languages like Malayalam, so the above scores are evaluated without using normalization. With normalization (for a fair comparison with other models on this platform), the results are instead:

WER: 11.49

This Colab can be used as a starting point to further finetune the model.

Usage instructions

Given an audio sample audio (this can be anything from a numpy array to a filepath), the following code generates transcriptions:

python

from transformers import pipeline, WhisperProcessor

processor = WhisperProcessor.from_pretrained("thennal/whisper-medium-ml")
forced_decoder_ids = processor.get_decoder_prompt_ids(language="ml", task="transcribe")
asr = pipeline(
        "automatic-speech-recognition", model="thennal/whisper-medium-ml", device=0,
    )
transcription = asr(audio, chunk_length_s=30, max_new_tokens=448, return_timestamps=False,  generate_kwargs={
        "forced_decoder_ids": forced_decoder_ids, 
        "do_sample": True,
    })

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 32
eval_batch_size: 16
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
training_steps: 8000
mixed_precision_training: Native AMP

Framework versions

Transformers 4.26.0.dev0
Pytorch 1.13.0+cu117
Datasets 2.7.1.dev0
Tokenizers 0.13.2