5×reduction in word error rate after fine-tuning
The challenge
What we walked into.
A Rabbi needed accurate transcription and summarization of sermons that move fluidly between Hebrew and English. Off-the-shelf speech-to-text tools failed on the code-switching and lost the cultural and contextual meaning.
What we built
The solution.
01
Fine-tuned OpenAI's Whisper speech-to-text model on more than 10 hours of custom bilingual audio.
02
An AI pipeline that turns spoken sermons into clear, structured English summaries while preserving context.
The results
What changed.
01
Roughly 5× lower word error rate and 8× lower character error rate versus the base model.
02
Sermons become concise, culturally faithful English summaries — automatically.