VoxFlow Pro: free push-to-talk Whisper dictation for macOS (Neural Engine to GPU fallback, stall watchdog)

[pittner]

VoxFlow Pro is a free push-to-talk dictation app for macOS. Hold Globe (Fn), Caps Lock or right Shift, speak, release, and the transcript is typed into the app that has focus. It runs Whisper models locally through WhisperKit (Core ML). The model (default small) is downloaded once on first launch; after that transcription works offline.

• Universal binary (Intel + Apple Silicon), macOS 14.0+
• Developer ID signed and notarized, no account, no subscription
• Download + SHA-256 checksum: https://bemooore.com/voxflow/mac/?utm_source=github&utm_medium=discussion&utm_campaign=whisper_showtell

Two things from running Whisper on Apple hardware that may help anyone doing the same. Both are in our development code; the public 1.0 download is from May and does not have them yet.

1. On some macOS versions the Neural Engine refuses large-v3, so keep a GPU fallback.
On macOS 26.5 with large-v3, Core ML logged E5RT errors on the Neural Engine and the transcription threw, three attempts in a row, so the recording was lost. We now load with the audio encoder on .all and the text decoder on .cpuAndNeuralEngine. If the load or an inference throws, we reload the same model once with encoder and decoder on .cpuAndGPU and run the same audio again. The fallback stays on for the session. A GPU load starts slower, but it returns text.

2. A transcription can hang without an error, so use a watchdog instead of a fixed timeout.
Some runs simply stopped reporting progress. We now cancel a run that reports no progress for max(60 s, 2 x audio length), reload the model and try the same audio once more. If the retry also stalls, the user gets an explicit error, never an empty transcript. Progress comes from the timestamps of the tokens decoded so far, because "window index x 30 s" jumped to 95 % on any clip shorter than 30 s.

Decoding settings, in case they are useful for comparison: temperature 0.0 with +0.2 fallback up to 5 times, VAD chunking (skips silence, which helps long voice memos), language detection only when the user has not set a language.

Feedback on accuracy outside English is welcome; we test mostly Slovak and English. Thanks to OpenAI for releasing Whisper under the MIT license.