Yapflow

Yapflow is two pieces on the same home network: a small Mac app, and a server on a Jetson Orin Nano sitting on my desk. Hold Right-Command, say the sentence, let go — a second later the cleaned-up text is typed into whatever window was in front, whether that’s a terminal, a browser, or a text field in an app that knows nothing about any of this.

The reason it exists is the last clause of that sentence. Good dictation is a solved problem if you are willing to stream your microphone to somebody else’s API all day. I wasn’t, so the model runs on a board I can unplug.

It’s a personal tool, not a product — there is no installer and no account, just a repo and a systemd unit. Source is on github.com/shreyas-sreedhar/yapflow.

• Hold to talk: The Mac app captures the microphone while the hotkey is down, resamples it, and streams it to the Jetson over a single WebSocket that stays open between dictations rather than reconnecting each time.
• Transcribe while you speak: Moonshine v2 runs on the Jetson and emits partial transcripts as the audio arrives, so the final text is ready almost as soon as the key comes up rather than starting from scratch on release.
• Tidy the result: A deterministic pass removes filler words and applies a personal dictionary — the names and jargon you have corrected before, which generic speech models reliably get wrong.
• Put it at the cursor: A small Swift helper writes to the clipboard and synthesizes Cmd+V. Unglamorous, but it is the one method that works in every macOS app instead of only the well-behaved ones.

• Curate for the hardware: Combined FLEURS and quality-filtered IndicVoices into 7,185 valid Kannada clips: 19.46 hours of mono, 16 kHz audio. Clips over 25 seconds, empty transcripts, persistent noise and unclear speech were removed before training.
• Make evaluation reproducible: Produced a deterministic 80/10/10 manifest split: 5,748 training samples, 718 validation samples and 719 held-out test samples.
• Try the memory-efficient path: Quantized the base model and applied LoRA so adaptation could fit the board instead of requiring a separate training machine.
• Find the actual bottleneck: The first run did not learn Indic speech. The base tokenizer had no useful Kannada vocabulary, while standard LoRA trained attention projections and left the newly needed embedding rows frozen.