Voice Messages

lynox accepts voice input on two surfaces:

• Web UI (primary) — record in the browser, transcript is sent back for review or submitted directly as a message.
• Mail / Inbox — forward a voice attachment (m4a / mp3 / ogg) to your connected mailbox and lynox transcribes it before running the task.

Transcription happens server-side. Two providers are supported; you can choose which one runs.

Keyboard shortcut

In the Web UI, double-tap ⌘ (macOS) or Ctrl (Windows/Linux) to start or stop recording — no chord, just two quick taps on the bare modifier within 350 ms. The shortcut is intentionally collision-free with every other browser/OS binding (a bare modifier is never used as a hotkey anywhere else), so it works the same in any focused field. Stop with the same gesture or click the microphone icon.

Provider matrix

Provider	Speed (60 s clip)	German WER (business speech)	Cost	Hosting	When to use
Mistral Voxtral	~2 s	~10 % on mixed DE/EN	$0.003/min	Mistral La Plateforme (Paris)	Default for cloud and BYOK setups. Fast, EU-hosted, no training on your audio.
whisper.cpp	several seconds (CPU)	~23 % on mixed DE/EN	free	Local (your server)	Air-gapped self-host. OSS fallback when no Mistral key is set.

Numbers come from the Phase 0 spike on ten self-recorded German-business-speech clips (see PRD voice-transcription-v2.md). whisper.cpp uses the ggml-base model on CPU.

Configuration

Provider selection

Set transcription_provider in your ~/.lynox/config.json:

{
  "transcription_provider": "auto"
}

Values:

• "auto" (default) — use Mistral if MISTRAL_API_KEY is set, otherwise whisper.cpp.
• "mistral" — force Mistral Voxtral. Transcription fails if the key is missing.
• "whisper" — force local whisper.cpp. Transcription fails if the binary/model are missing.

The LYNOX_TRANSCRIBE_PROVIDER environment variable overrides the config value.

Mistral Voxtral

Requires an API key from console.mistral.ai:

export MISTRAL_API_KEY=...

lynox calls the /v1/audio/transcriptions endpoint with model voxtral-mini-2602. Only the documented parameters are sent (file, model, language). Your recording is transmitted to Mistral’s EU infrastructure; per Mistral’s terms, customer audio is not used to train models.

whisper.cpp

Requires the whisper-cli binary and a ggml model on the host:

# macOS
brew install whisper-cpp ffmpeg
# ggml model
curl -L -o ~/.local/share/whisper/ggml-base.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin

Defaults to the base model; falls back to tiny for clips under 10 seconds.

Glossary repair

Speech-to-text systems mis-hear proper nouns and product vocabulary in predictable ways. lynox applies a two-layer glossary to the raw transcript before returning it:

• Core glossary — lynox product vocabulary (Setup Wizard, Go-Live, Knowledge Graph, …). Seeded from known mishearings; extended via PR as new surface-area names ship.
• Session glossary — built at call time from your own context: CRM contact names, registered API/tool names, recent thread titles, Knowledge-Graph entity labels, custom workflow names. A short edit-distance match (≤2) rewrites nearby tokens when they are not in a common-language stop list — so Rolland becomes Roland but rund is left alone.

Both passes run in single-digit milliseconds. The glossary never leaves the lynox process; your vocabulary is never part of the audio-transcription API request.

Privacy

• Mistral-hosted audio: sent to Paris, not retained for training, not stored post-transcription per Mistral’s terms.
• whisper.cpp: audio never leaves the server; transient /tmp files are deleted after each transcription.
• The Web UI shows a short privacy hint under the voice button indicating which provider is in use.
• No audio is retained by lynox — only the final text goes into the thread history.

Troubleshooting

• “Transcription not available” — check that MISTRAL_API_KEY is set or that whisper.cpp is installed.
• Short clips mis-transcribed — whisper.cpp uses the tiny model for clips ≤10 s; record a slightly longer utterance to get the more accurate base model.
• Product term still wrong — if a new lynox product name ships and is mis-heard, add it to the core glossary (src/core/transcribe/glossary/core-terms.ts). Contact/tool names self-heal from the session glossary.