Voice chat¶

gmlx talk is a voice client for the models that the gmlx server serves. It listens for a wake phrase, sends what you say to the model and speaks the reply as it streams. You can interrupt a reply, change the client's settings and run the client from a script.

gmlx talk                                # the default model; say "hey assistant", then speak
gmlx talk qwen3.8-27b-ud-q6 --mode ptt   # a named model; press Space to speak

Speech recognition, speech synthesis and the model all run in the server, so the server needs its speech services first, as Setting up describes. With a remote server, that machine does the work while the microphone and speakers stay on your Mac.

Setting up¶

Install the talk extra, which includes the audio, wake phrase and speech packages. A Homebrew install already has it, and Optional features shows how to add it to a uv or pip install.

Turn on the server's speech services:

server:
  stt: whisper-turbo
  tts: kokoro

Run gmlx init to choose the voice, the wake phrase, the listening mode and the menu bar app's hotkey. The wizard asks when it sets up both speech services, and offers to install a missing talk extra.
Run gmlx restart so that the server starts the services, and then run gmlx talk.

A served model that hears audio, such as Nemotron 3.5 Transcribe, can hear you and answer you in one model. The server then needs only tts, because server.stt left unset picks that model.

The first session downloads two small models, for the wake phrase and for speech detection, into the cache folder that Where files are on disk lists.

macOS asks once for permission to use the microphone, in your terminal's name. The mic never works in talk shows how to allow it later.

A session¶

gmlx talk starts the server if it is down, and then asks it to keep the model loaded until the session ends. With no model named, it uses talk.model, then the server's default model, then the only model that the server serves. The first line names the model, the voice and the mode:

[talk] qwen3.8-27b-ud-q6, voice default, mode wake, say "hey assistant"  (q quits, / commands)
you: what's a good name for a gray cat?
  How about Ash?
  It suits a gray coat, and it is short enough for the cat to learn.

Each thing you say prints after you:, and each part of the reply prints indented as it is spoken. A status line at the bottom shows whether the session is idle, listening, capturing, transcribing, thinking, speaking or muted. A rising chime means that the microphone is open, and a falling chime means that it has closed.

To type a message instead, start typing. The first key opens a line editor, Enter sends the line, and the reply is spoken as usual. A line that you type while the model is still answering is dropped.

Listening modes¶

The mode sets how a turn starts:

Mode	How a turn starts
`wake`	You say the wake phrase, and then speak. The default.
`vad`	Any speech starts a turn.
`ptt`	Space opens the microphone. The turn ends after a pause, or when you press Space again.
`text`	You type every message, and the replies are spoken. Needs only the text-to-speech service.

Any English text works as the wake phrase, with no training needed. The default is "hey assistant", and /wake changes it during a session. Choose a phrase that the model is unlikely to say, because the client also hears the speakers.

An utterance ends after 550 milliseconds of silence by default, which talk.vad.silence_ms changes. In ptt mode, pressing Space before you speak closes the microphone again.

A clip of 1.5 seconds or less that transcribes to a common filler, such as "okay", "thanks" or "bye", is dropped, because speech recognition often hears these in silence. Say a longer phrase, such as "okay, go on".

Keys and commands¶

These keys work while the status line shows:

Key	Effect
Space or Esc	Stops the reply while the model transcribes, thinks or speaks. In `ptt` mode, Space also opens and closes the microphone.
`m`	Mutes or unmutes the microphone, at any time.
`q` or Ctrl-D	Quits the client when no reply is in progress.
Any other printable key	Opens the line editor, starting the line with that character.

Because q, m and Space act at once, a typed message cannot start with them. A line that starts with / runs a command:

Command	Effect
`/voice [name]`	Lists the server's voices, or switches to one.
`/speed <number>`	Sets the speech speed, from 0.25 to 4.
`/mode wake\\|vad\\|ptt\\|text`	Switches the listening mode.
`/wake [phrase]`	Shows or changes the wake phrase.
`/mute`	Mutes or unmutes the microphone.
`/system [text]`	Sets the system prompt, or clears it with no text, and starts the conversation again.
`/reset`	Starts the conversation again.
`/memory`	Shows and manages the assistant's memory.
`/devices`	Lists the audio devices.
`/help`	Lists the commands.
`/quit`, `/exit`, `/q`	Quits the client.

Interrupting a reply¶

Space or Esc stops the speech within about 150 milliseconds and cancels the turn. In wake mode, the wake phrase also interrupts, because the client keeps listening for it while the model answers. The microphone then opens for your next turn. If you say a stop phrase such as "stop", "cancel" or "never mind" instead, the client goes back to waiting for the wake phrase.

The vad and ptt modes do not listen while the model answers, because the microphone would pick up the reply and transcribe it. In these modes, interrupt from the keyboard.

The assistant by voice¶

With talk.brain: assistant, the assistant answers instead of the plain model. It can call tools from MCP servers during a turn, and it can remember facts across sessions. This example adds two MCP servers that need no API keys, and turns on memory:

server:
  stt: whisper-turbo
  tts: kokoro
  embeddings: qwen3-embed-0.6b     # Required for memory.
  rerank: qwen3-rerank-0.6b        # Optional. Reorders recalled memories.

talk:
  model: qwen3.8-27b-ud-q6@instruct
  brain: assistant

assistant:
  mcp:
    - name: files
      command: [npx, -y, "@modelcontextprotocol/server-filesystem", "/Users/me/notes"]
    - name: web
      command: [uvx, mcp-server-fetch]
  memory:
    enabled: true

The filesystem server needs Node, and the fetch server needs uv. Assistant covers choosing a model, the tools and the memory store. Decisions as a tool adds a tool that answers questions through /v1/systemone, and Granite Switch can hear, answer and decide in one model.

During a turn, the status line names each tool as the assistant uses it, and only the answer is spoken. After the turn, the assistant stores what it learned in the background, so memory does not delay the spoken reply. When you interrupt a tool call, the conversation keeps what you heard and drops the unfinished call.

Settings¶

The talk block of the configuration file sets the model, the voice, the mode, the wake phrase and the speech detection thresholds. Most of its keys have a flag under gmlx talk, and a flag wins over the file. The menu bar app runs the same session without a terminal and adds a tap-to-talk hotkey.

To shorten the wait for a reply, change these settings:

Setting	Effect
A lower `talk.vad.silence_ms`, such as 400	The turn ends sooner after you stop, and a pause in a sentence ends it more often.
`stt: whisper-turbo-q4`	Transcription is faster.
A lower `talk.max_tokens`, such as 512	Replies are shorter.

A remote server and scripts¶

--base-url http://host:8080/v1 points the client at another server, and --api-key gives its key. Speech recognition and synthesis then run on that machine. Without --base-url, the client uses your configured server and starts it when it is down, unless you pass --no-start.

--once answers one utterance and exits, with no wake phrase, for scripts and for checking a setup:

gmlx talk --once