Fine-Tune Whisper for Yiddish Lectures

I have a growing collection of Yiddish lecture recordings—some streamed live, others stored as audio files—that must be transcribed with near-publishing accuracy. Whisper’s out-of-the-box performance is promising, yet it still drops domain-specific vocabulary and speaker nuances. I want a fine-tuned Whisper model that reliably converts these lectures to text, whether the audio arrives in real time or from an archive.

You will receive aligned Yiddish audio–text pairs plus a small evaluation script. Use whichever Whisper implementation you prefer—OpenAI’s official repo, Hugging Face Transformers, or another PyTorch-based fork—so long as the final system:

• archives single-digit word-error-rate on my hold-out set
• preserves accurate timestamps for downstream subtitling
• runs on a single A100 GPU and can be quantized for CPU inference



Deliverables I need back:

• the fine-tuned model checkpoint(s)
• reproducible training and inference scripts or notebooks, with hyper-parameters clearly logged
• a brief report comparing baseline vs. tuned performance (WER, CER, latency)

Once these items pass my internal test—live Zoom feed and batch file transcription—the project is complete.