p99 labs
Research Note · 2026

On-Device Speech Systems

speech recognition · local inference · Apple Silicon

Abstract. P99 Labs develops speech systems that run entirely on local hardware. Performing recognition and post-processing on the device removes the network round-trip that defines cloud dictation: no audio is transmitted, no account is required, and no data is retained off-device. This note states our design principles and describes our first system, Blurt, a faithful voice keyboard for macOS.


Local inference

All recognition runs on the user's machine. Privacy is a property of the architecture rather than of policy: data that never leaves the device cannot be logged, retained, or disclosed. We treat this as a constraint on every system we ship, not an optional mode.

Faithful transcription

Our systems reproduce what was said. We normalize punctuation, spacing, and the orthography of known terms; we do not paraphrase, summarize, or generate. Where generative rewriting cannot be verified as correct, we omit it.

Efficiency

Useful speech models need not be large. We target compact models with sub-second latency on commodity Apple Silicon, operating fully offline after a one-time download.

Systems

Blurt — a voice keyboard for macOS. Push-to-talk dictation that inserts recognized text at the cursor in any application, multilingual, entirely on-device.

P99 Labs · correspondence: research@p99lab.com