Whisper Large V3 — Hindi Fine-tuned 🎙️
Fine-tuned version of openai/whisper-large-v3 for Hindi speech recognition using LoRA (PEFT).
Model Details
| Parameter | Value |
|---|---|
| Base Model | openai/whisper-large-v3 |
| Dataset | SPRINGLab/IndicVoices-R_Hindi |
| Train Samples | ~14,800 |
| Eval Samples | ~740 |
| Training Steps | 4400 |
| Best Val Loss | 0.1167 |
| LoRA Rank | 8 |
| LoRA Alpha | 16 |
| Baseline WER | 32.10% |
| Fine-tuned WER | 28.98% |
| Method | LoRA fine-tuning (PEFT) |
| Precision | bfloat16 |
Usage
from transformers import pipeline
asr = pipeline(
task='automatic-speech-recognition',
model='Sa1Krishna/sema-whisper-large-v3-hindi-finetuned',
chunk_length_s=30,
device=0
)
result = asr(
'hindi_audio.wav',
generate_kwargs={
'language': 'hindi',
'task': 'transcribe'
}
)
print(result['text'])
Training Details
Trained on SPRINGLab/IndicVoices-R_Hindi — a large scale Hindi speech dataset with diverse speakers.
Training Config
- Optimizer : AdamW
- LR : 1e-4
- Batch size : 16 (effective)
- Precision : bfloat16
- Framework : HuggingFace Transformers + PEFT
Limitations
- Optimised for Hindi speech only
- May struggle with heavy accents
- Short utterances work best