Butter

Why It Won — our reading

At a Whisper hackathon full of transcription demos, Butter chose not the cleanest audio but the hardest speakers. Stuttered speech is where ordinary recognition tends to break down, so getting usable text out of it is about the strongest showcase a speech model could ask for. Framing the work around a person's difficulty rather than a capability, and answering questions about stuttering instead of only transcribing, is likely what carried it into the top three.