whisper-tiny-az

openai/whisper-tiny (39M parameters) fine-tuned for Azerbaijani speech recognition.

Azərbaycan dili üçün öyrədilmiş ən kiçik Whisper modeli — sürətli, zəif cihazlarda da işləyir.

Results (FLEURS az_az test)

Text normalized before scoring: Azerbaijani-aware lowercase (I→ı, İ→i), punctuation removed.

Usage

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="Kenan321/whisper-tiny-az", chunk_length_s=30)
print(asr("audio.wav", generate_kwargs={"language": "az", "task": "transcribe"})["text"])

Training

• Data: FLEURS az_az train + validation (~3,000 utterances, read speech). The test split was not used for training.
• 3,000 steps, batch size 32, learning rate 1e-4 (linear, 200 warmup), fp16, SpecAugment (time/feature mask 0.05).
• Hardware: single RTX 5070 Laptop GPU (~90 min).

Limitations

• Small model trained on a small dataset of clean read speech: expect higher error on noisy audio, spontaneous speech and dialects.
• Long recordings (>30 s) are processed in chunks and may lose words at chunk boundaries.
• For best accuracy use a larger model (e.g. a fine-tuned whisper-small or whisper-large-v3-turbo).