Voice Input and Read Aloud
LanMate supports several voice workflows: voice wake-up (say a wake word to start talking), hold-to-talk (speech to text), voice conversation (continuous dialogue with playback), and read-aloud (text to speech). With local engines, transcribed and synthesized text never leaves your machine; when a cloud endpoint is configured (FunASR / OpenAI-compatible transcription endpoints, MiniMax / Edge synthesis, etc.), audio or text is uploaded to the corresponding service, subject to that service’s privacy policy.
Voice Wake-Up
Enable it under Settings → Voice Transcription & Synthesis → “Voice Wake-Up”, then speak the wake word to start a voice conversation:
- The default wake word is “Xiao Lan Xiao Lan” (小蓝小蓝) and can be customized (up to 30 characters)
- The wake model ships with the installer (about 3.3MB); recognition happens on your machine, no internet needed
- The “Test” button next to the wake-word input opens a dialog: live waveform plus a 10-second one-shot wake detection to confirm your phrase is reliably recognized
- After waking, the voice-conversation panel opens automatically and starts listening; wake listening pauses while the panel is open
Great for hands-off-the-keyboard moments: walk into the office, call out, and start assigning work.
Hold to Talk
The microphone button on the right of the input: hold to speak, release to transcribe:
- When the input box is empty you can also long-press the input box to start talking right away; when there’s already text, tap the microphone button to start recording and tap it again to stop
- The transcript can be sent directly, or filled into the input for editing first (configurable in settings)
- Too-short clips are rejected — no empty messages
- Requires a transcription engine configured under Settings → Voice Transcription & Synthesis
Great for when your hands are busy, or when speaking is faster than typing.
Voice Conversation
The “Voice Conversation” button on the input enters a full-screen voice mode:
- Speak; a detected pause sends automatically
- Agent status shows while thinking
- Replies are spoken automatically
- Speak during playback to interrupt — no need to wait
The whole thing flows like a phone call. Live captions show everything you said and every reply on screen.
Read Aloud
The read-aloud button on a message speaks the text. Settings offer:
- Auto-read: off / final replies only (default) / read the whole execution process
- Voice & speed: each engine has its own voice slots with previews; speed 0.5×–2.0×
- Read-aloud text automatically skips code blocks, images, and links — body text only
Engine Setup
Settings → Voice Transcription & Synthesis:
| Engine | Used for | Notes |
|---|---|---|
| Transcription model | Hold-to-talk, voice attachments | SiliconFlow (SenseVoiceSmall), Groq (Whisper Turbo), a FunASR endpoint, an OpenAI-compatible custom endpoint, local Whisper, or local Qwen3-ASR |
| Real-time transcription | Voice conversation | Local streaming engine, or a “cloud transcription endpoint” (FunASR / OpenAI-compatible): whole sentences are uploaded for recognition and captions refresh per sentence |
| Synthesis model | Read aloud | Local engines (MeloTTS / Kokoro / CosyVoice2), MiniMax cloud voices, Edge TTS (Microsoft online, free, no key needed), custom endpoints, or custom commands |
Additional notes:
- Local Whisper install: when unconfigured, the settings page shows an “Install” button that installs dependencies into the shared environment and streams the model download (auto-falls-back to a mirror on failure), with live progress
- Local synthesis engines stay resident: once installed, synthesis responds faster; the selected engine pre-warms and auto-releases when idle
- MiniMax requires an API key (China / international regions supported); Edge TTS is free and works out of the box
- With cloud engines, read-aloud text or audio is sent to the corresponding provider
Lansenger Voice Messages
Voice messages received via Lansenger channels are auto-transcribed and enter context with the attachment. Send voice from the channel; the desktop handles it.