whisper-small-he

Hebrew fine-tuned Whisper Small for automatic speech recognition.

Training

• Base model: openai/whisper-small
• Dataset: ivrit-ai/whisper-training (~400h Hebrew)
• Method: Supervised fine-tuning with Seq2SeqTrainer
• Steps: 5,000 (streaming, effective batch size 16)
• Hardware: Apple M4 (MPS), fp32
• Best eval WER: 0.368 (on 200-sample test split, step 4000)

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("amitkot/whisper-small-he")
model = WhisperForConditionalGeneration.from_pretrained("amitkot/whisper-small-he")

model.generation_config.language = "he"
model.generation_config.task = "transcribe"

Training pipeline

Trained using whisper-acft-pipeline:

uv run python scripts/finetune.py --config configs/hebrew_small_finetune.yaml

See also

• amitkot/whisper-small-he-acft — ACFT-optimized version of this model for short audio (FUTO Keyboard)
• amitkot/whisper-tiny-he — Smaller/faster variant