A solo developer has released Phonon, an open-source voice-typing app for Apple Silicon Macs. It aims to reproduce Wispr Flow’s core workflow: turning dictated speech into polished text suited to the destination app.

The release includes the refiner model, a LoRA fine-tune of LiquidAI’s LFM3-1.7B. Developers can inspect and run the complete pipeline without a hosted transcription API, although the model’s license limits commercial use by larger organizations.

Two local models, one dictation path

On first launch, Phonon downloads exact revisions of Parakeet TDT 0.6B v2 for speech recognition and the 1.7-billion-parameter refiner for correction. Parakeet produces a raw transcript; the refiner then fixes likely misrecognitions, removes verbal filler, handles formatting, and adapts the text to the active app.

Both stages run on the Mac after the initial downloads. Dictation stays on-device, and the pipeline avoids network variability and API round trips. Pinning each model revision also makes installations more reproducible.

The 65 ms claim needs boundaries

The author reports key-up latency, the delay after releasing the dictation key, of 65 milliseconds for Phonon and about 520 milliseconds for Wispr Flow on the same clips and M4 Max hardware.

Phonon uses a Rust backend and a native Metal engine. Each decoding step runs as a fused megakernel, combining work into a single GPU launch, while a DFlash2 speculative drafter predicts candidate tokens to reduce serial decoding work.