Transcription in under 50 milliseconds

Hearsy uses the Parakeet TDT v2 model, which is optimized specifically for Apple Silicon's Neural Engine. The model runs as a compiled CoreML graph, bypassing Python overhead entirely. Combined with streaming audio processing, the end-to-end latency from speech to text is typically under 200ms, with the transcription step itself completing in under 50ms.

Hearsy requires Apple Silicon (M1 or later). The Neural Engine acceleration that makes sub-50ms transcription possible is only available on Apple Silicon chips. Intel Macs don't have the Neural Engine hardware required for this level of performance.

The Parakeet TDT v2 model is approximately 450MB. It downloads once during initial setup and runs entirely on your Mac after that. The Whisper v3 Turbo model is a separate download of similar size if you need broader language support.

For typical dictation (up to a few minutes), the processing time is nearly constant because Parakeet uses streaming inference. Longer recordings may take slightly more time for final processing, but the perceived latency remains minimal because text appears progressively.

The Neural Engine is designed for efficiency. During active dictation, you'll see minimal battery impact — comparable to playing music. The model only runs while you're actively recording, and idle power consumption is effectively zero since Hearsy sits quietly in your menu bar.

Yes. Whisper v3 Turbo typically takes 1-2 seconds for transcription compared to Parakeet's sub-200ms. The tradeoff is that Whisper supports 99+ languages with automatic detection, while Parakeet covers 25 languages. You can switch between engines instantly based on your needs.