Views
No views yet
openai/whisper-small, fine-tuned for Hindi ASR on 500 FLEURS (hi_in) utterances. ~35 minutes on a free Colab T4.| Voice | Baseline WER | LoRA WER | Relative Δ | Sarvam Saaras v3 |
|---|---|---|---|---|
| Hindi Female (n=30) | 0.515 | 0.362 | −29.8% | 0.237 |
1from transformers import WhisperForConditionalGeneration, WhisperProcessor
2from peft import PeftModel
3
4processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")
5model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
6model = PeftModel.from_pretrained(model, "manishehehe/whisper-small-hindi-lora")
7model.eval()
8
9ids = model.generate(input_features=input_features, language="hi", task="transcribe", max_new_tokens=225)
10print(processor.tokenizer.batch_decode(ids, skip_special_tokens=True)[0])q_proj/v_proj (r=32, α=64, dropout=0.05) — ~1.8M trainable params (0.7% of base). 500 steps, effective batch 32, LR 1e-3, fp16, single T4.