AyikuBen/whisper-ga-finetuned-demo

Whisper-tiny · Ga (fine-tuned demo)

Short summary
This is a small demo model: openai/whisper-tiny fine-tuned on ~200 Ga clips (16 kHz) to validate an end-to-end ASR pipeline. It’s for learning/experimentation, not production.

🔧 Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import soundfile as sf, torch
MODEL_ID = "AyikuBen/whisper-ga-finetuned-demo"
proc = WhisperProcessor.from_pretrained(MODEL_ID)
mdl  = WhisperForConditionalGeneration.from_pretrained(MODEL_ID)
mdl.config.forced_decoder_ids = None
mdl.config.suppress_tokens = []
a, sr = sf.read("example.wav")
ip = proc.feature_extractor(a, sampling_rate=sr, return_tensors="pt")
ip = {k: v.to(mdl.device) for k, v in ip.items()}
with torch.no_grad():
    ids = mdl.generate(ip["input_features"], max_length=225, do_sample=False, no_repeat_ngram_size=3)
print(proc.tokenizer.batch_decode(ids, skip_special_tokens=True)[0])