voice

Local speech-to-text and read-aloud on the iii bus. Nothing leaves the machine by default: a small streaming recognizer produces live text, while bundled Parakeet or a configured host whisper-cli can produce accurate final text. An OpenAI-compatible endpoint or any speech provider registered with llm-router can replace either half through configuration.

When to Use

• A user attached or named a WAV recording and wants its text: voice::transcribe.
• A surface streams microphone audio and wants live text: voice::dictation::*.
• A reply should be spoken: voice::speak.

Boundaries

• Audio in is 16 kHz mono 16-bit PCM for dictation; WAV files of any rate for voice::transcribe. No other container is decoded.
• The bundled models are English. Portuguese and other languages need stt.backend: whisper_cpp with a multilingual GGML model, or a remote backend.
• For natural local read-aloud, install piper-tts, enter your language in Voice's Piper language filter, choose a matching voice from the bundled multilingual catalog, explicitly download it, and select tts.backend: piper. piper-pt-br-faber-medium remains available for Brazilian Portuguese. Filtering does not download or change the active voice. tts.piper.command is an executable path, not a shell command. No fallback to robotic eSpeak occurs if Piper is unavailable.
• tts.piper.device defaults to auto (request NVIDIA CUDA, allow ONNX Runtime CPU fallback; retry once on CPU if the GPU-preferred process fails). cpu forces CPU. Auto allocates 30 seconds per attempt; CPU-only gets 60 seconds. GPU requires compatible onnxruntime-gpu/CUDA/cuDNN in Piper's own environment. No drivers are auto-installed; doctor reports policy, not verified GPU usage.
• In Console, Read selection speaks only highlighted text in that chat. Voice chat is opt-in per browser/conversation and reads live assistant text in sentence/phrase chunks before the turn ends (including visible progress, never thinking/tool arguments). It deduplicates snapshots and flushes only the unread tail at completion; it does not send messages or keep the microphone open. Stop disables automatic reading.
• voice::speak accepts text_format: markdown (default) to read prose without Markdown markers, hidden link URLs or block code; plain preserves literal text. Chat selections use plain because they are already rendered. The character cap applies after conversion. No messages are modified and no remote links are fetched.
• Every voice::speak backend returns audio_base64 and mime to the caller. host generates WAV using say/espeak in file-output mode, never server playback. The requesting browser plays/stops its own audio; other clients are unaffected.
• Dictation sessions idle past session_idle_secs are closed by the worker.

Functions

• voice::transcribe — a WAV file (path or base64) to text with timestamped segments.
• voice::dictation::start — open a session; transcript events go to output_function_id.
• voice::dictation::push — feed one base64 PCM chunk (rising seq).
• voice::dictation::stop — close a session and return its transcript (discard to drop it).
• voice::dictation::list — open sessions.
• voice::speak — read text aloud; returns a speech_id.
• voice::speech::prepare — stateless Markdown snapshot preparation (text, complete); returns plain text and max_chunk_chars. Withhold unfinished syntax, then queue plain-text voice::speak clips in the requesting browser.
• voice::speak::stop — legacy compatibility, returns stopped: 0. Stop browser audio locally.
• voice::models::list — local models, whether each is installed, and each one's license, author and source (Parakeet TDT 0.6B v2 is CC BY 4.0: keep the attribution if you copy its files elsewhere).
• voice::models::download — install a local model, checksum-verified.
• voice::models::remove — delete a downloaded model.
• voice::doctor — backends, model state, open sessions.

Reactive triggers

• voice::transcript — partial, final, closed and error events of dictation sessions (filter session_id).
• voice::session-started, voice::session-stopped — session lifecycle.
• voice::model-progress — download progress, one event per megabyte and a final done.