Overview
Voice mode combines:- Speech recognition — Converts your voice to text
- AI chat — Processes your message and generates a response
- Text-to-speech — Speaks the response back to you
Getting Started
Open Voice in the sidebar. The first time, Izwi lists the models voice needs and downloads them with one click. Models load when you start a conversation, so there is no separate load step. Voice has two engines, chosen in Voice settings › Engine:
To download them from the command line instead:
Using Voice Mode
- Navigate to Voice in the sidebar
- Click Start talking (or press
Space) and allow microphone access if asked - Speak naturally; Izwi answers when you pause
- Start talking while Izwi speaks to interrupt it
- Click End (or press
Space) when you are done
Controls
Voice Settings
- Voice — Choose the assistant’s voice; press play to hear a preview
- Instructions — How the assistant should answer. Saved when you leave the field and used from your next conversation; Reset to default restores the original
- Memory — When on, Izwi remembers lasting preferences and facts you mention. Manage memories lists them so you can forget one or all
- Engine — Standard or Compact, the status of its models, and a button to load or unload them
API Routes
Voice mode uses preview APIs for the voice profile, observations, saved sessions, and low-latency WebSocket transport. See the API Reference for persisted voice state and Voice Realtime for the WebSocket protocol.LFM3-Audio-2B-GGUF answers spoken turns with interleaved text and speech,
through the same speech-to-speech routes as LFM2.5-Audio-1.5B-GGUF: name
either as s2s_model_id in the WebSocket’s unified mode. The Compact
engine uses LFM3-Audio-2B-GGUF by default, which picks its own voice in
conversation. LFM2.5 Audio stays selectable in Voice settings › Engine,
with its four voices, and with only LFM2.5 Audio downloaded the Compact engine
uses it.
A spoken turn can call tools when its model can: a chat model that calls tools
(Qwen3, Qwen3.5, LFM2.5, LFM3) in the modular pipeline, or LFM3 Audio. It may
call the server’s built-in tools (the local date and time; on by default with
LFM3 Audio, on request in the modular pipeline), function tools a WebSocket
client declares in input_stream_start and answers with
function_call_output, and the tools of MCP servers the session names. The reply calls a tool, reads its result, then
speaks the answer. See
Voice Realtime for the fields and events.
Tools, tool servers, agents that hand over to each other, and the switches
for how Izwi takes turns are set in Voice settings. The
Voice agents guide walks through them.
Tips for Best Results
- Use a good microphone — Built-in laptop mics work but external mics are better
- Minimize background noise — Find a quiet environment
- Speak clearly — Natural pace, clear pronunciation
- Wait for responses — Let the AI finish before speaking again
Keyboard Shortcuts
Shortcuts are ignored while focus is inside a text field or on a button.
Troubleshooting
No audio input detected
- Check your microphone permissions in system settings
- Ensure the correct input device is selected
- Test your microphone in another application
”Couldn’t load the voice models”
The models do not fit in the memory Izwi can use alongside the other models it has loaded. Unload models you are not using on the Models page, or switch to the Compact engine in Voice settings › Engine.Responses are slow
- The first reply after starting includes loading the models; later replies are faster
- Try the Compact engine
- Check system resources (RAM, CPU usage)
Poor transcription accuracy
- Speak more clearly and slowly
- Reduce background noise
- Use an external microphone
See Also
- Chat — Text-based conversations
- Transcription — Batch audio transcription
- Text-to-Speech — Generate speech from text