speech-to-text

Whisper's prompt parameter has a hidden 224-token limit. Overflow is cut from the front without warning, which can turn keyword biasing into a bug.

Jul 28, 2026 9 min read whisper speech-to-text prompt-engineering tokenization go

We ranked a domain vocabulary by corpus frequency and the top 20 came back all generic words. How a planned data gate kept a bad axis from shipping.

Jul 28, 2026 8 min read ranking corpus-frequency speech-to-text data-quality go