laasyamb/whisper-ta

Whisper Small Tamil

This model is a fine-tuned version of openai/whisper-small on the Common Voice 24.0 - Tamil dataset.

Model description

The model converts spoken Tamil audio into written Tamil text. It was fine-tuned using the Mozilla Common Voice 24.0 Tamil dataset.

• Base model: openai/whisper-small
• Language: Tamil (ta)
• Task: Speech-to-text (transcription)

Intended uses & limitations

More information needed

Training and evaluation data

The model was fine-tuned on:

• Dataset: Mozilla Common Voice 24.0 – Tamil
• Type: Read, crowd-sourced speech
• Audio: 16 kHz mono
• Text: Tamil transcriptions
• Splits: Train / validation Common Voice contains speech from a diverse set of speakers, but may still include demographic and accent imbalances. It skews toward younger male speakers.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Final Training Loss = 0.130500
CER (Character Error Rate) = 0.4946198117007057

Framework versions

• Transformers 4.52.0
• Pytorch 2.9.0+cu126
• Datasets 4.4.2
• Tokenizers 0.21.4