Flow

Good dictation lives behind a subscription. Wispr Flow is $15/mo, and every word you speak goes through someone else's cloud. Flow replicates that experience entirely on the machine: hold a global hotkey, speak, release, and formatted text lands in whatever app has focus.

It's built by composing open-source inference offline. whisper.cpp on Metal handles transcription, a local Ollama pass cleans the text, and the result is injected without any per-app integration. The whole thing is free and works with no network.

Steps 01–03 run while the hotkey is held: capture audio and transcribe it. The moment the key is released, steps 04–06 fire: an LLM formatting pass, clipboard injection into the focused app, then restoring the prior clipboard.

• /01

  Clipboard injection, not the Accessibility API

  Text lands via clipboard plus a synthesized ⌘V, and the previous clipboard is restored after. It's the only approach that works across native apps, Electron, web views and terminals without wiring per-app accessibility hooks.

• /02

  Warm Whisper, sub-second after the first run

  A WhisperCache holds the Metal-initialized model between dictations. Only the first dictation after launch pays the model-load cost; every call after that returns in under a second on M-series.

• /03

  Hot path never touches SQLite

  Dictionary, snippets and style presets are denormalized into an in-memory PipelineConfig that rebuilds on each DB mutation. The dictation path reads the snapshot, so a write to the management UI never slows a dictation.

• /04

  Prompt regression suite from source

  A test suite extracts the system prompt straight out of the Rust source and asserts against it, so a casual edit to the formatting prompt can't silently drift the model's behavior.

• Q/01

  Why clipboard injection over the Accessibility API?

  +

  The Accessibility API needs per-app hooks and breaks inside Electron apps, web views and terminals. Clipboard plus a synthesized ⌘V works everywhere a paste works, which is everywhere. Restoring the previous clipboard afterward keeps it invisible to the user.

• Q/02

  Why a local + cloud engine toggle instead of picking one?

  +

  Local whisper.cpp and Ollama mean zero cost and nothing leaves the machine. Groq cloud Whisper is faster and sharper when the network is there. The toggle lets the user trade privacy and cost against speed and quality per their situation, rather than baking one choice in.

• Q/03

  Why cache the Whisper context between dictations?

  +

  Loading the Metal-initialized model is the expensive part. Holding it in a WhisperCache means that cost is paid once at first use. Subsequent dictations skip straight to inference and return in under a second, which is what makes hold-and-speak feel instant.

• Q/04

  Why an in-memory config snapshot instead of reading SQLite on the hot path?

  +

  Dictation latency is the product. Querying SQLite for dictionary, snippets and style on every dictation would put disk I/O in the critical path. Denormalizing into a PipelineConfig rebuilt on each mutation keeps writes correct and keeps the dictation path purely in memory.

• /01

  Move API keys into the macOS Keychain. Cloud engine keys currently sit in plain config; Keychain-backed storage is the right home for a credential on a desktop app, and it's a small migration done early.

• /02

  Switch to streaming transcription instead of record-then-transcribe. Today the audio is captured fully, then transcribed. Streaming the audio into Whisper as it's spoken would cut the felt latency on longer dictations close to zero.