Align text transcriptions in speech-to-text applications

This Blueprint enables you to align OpenAI’s Whisper speech‑to‑text models toward user-defined text. By supplying a custom list of phrases (e.g., brand names, technical terms, rare phrases), the model adjusts its transcriptions, improving accuracy for domain‑specific vocabulary, especially when you need reliable recognition of words that aren’t common in everyday language.

In the audio example below, you can compare the transcriptions before and after biasing the model with the text "Dileesh Pothan", which is the correct spelling of a name that does not appear often in the training data of the original model.

Without model alignment: "The rich potent as an Indian film director from Kerala who works in the Malayalam film industry."

With model alignment: "Dileesh Pothan is an Indian film director from Kerala who works in the Malayalam film industry."