whisper-tiny-he

Hebrew fine-tuned Whisper Tiny for automatic speech recognition.

Training

• Base model: openai/whisper-tiny
• 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
• Final eval WER: 0.659 (on 200-sample test split)

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("amitkot/whisper-tiny-he")
model = WhisperForConditionalGeneration.from_pretrained("amitkot/whisper-tiny-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_tiny_finetune.yaml

See also

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