Whistle: Speech to Text in 16.9 MB

Whistle is a new open-source speech recognition model that weighs just 16.9 MB and runs entirely on device CPUs without dependencies, supporting transcription in seven languages (English, German, French, Spanish, Italian, Dutch, and Polish) with the first token generated in 11 ms. The model integrates with the Needle CPU engine and provides transcription, word-level timestamps, and speech embeddings, with audio remaining private on the user's device.