Skip to content

title: "Talk to the assistant by voice" description: How a voice conversation works in the workspace chat: you speak, the assistant answers aloud and in the chat, what is not read out, which model it needs, and what to do when speaking stops or nothing is spoken.

Talk to the assistant by voice

What it does

In a voice conversation you speak a question and the assistant answers aloud. The answer is also stored in the chat as a normal message, so you can read it, search it and continue it by typing later. Sources, citations and generated files are shown in the chat and are not read out.

A voice conversation is switched on per workspace by the platform administrator. Where it is on, the Voice mode button appears in the composer next to the microphone button.

How to ask

  1. Open a conversation and click on the Voice mode button next to the microphone button. Allow the microphone if your browser asks.
  2. Speak your question and stop. After a short pause the assistant takes your words as one turn and starts to answer.
  3. Listen to the answer. You can speak the next question when Listening shows again; the conversation continues after every answer.
  4. Click on Stop answer to silence an answer early, or on End to leave the voice conversation. The next answer can take a moment longer after Stop answer.

Questions that were tested:

  • "Was ist der Unterschied zwischen Wetter und Klima?"
  • "Erkläre mir bitte kurz, wie die Fotosynthese funktioniert."
  • "Nenne mir drei Tipps für einen guten Schlaf."
  • "What is the difference between weather and climate?"
  • "Please explain in a few words how a rainbow forms."
  • "Give me three tips for a good night of sleep."

The assistant answers in the language you spoke (German, English or French). Spoken answers are short on purpose.

Tips

  • Speak in whole sentences and stop at the end of your question. A question that sounds unfinished waits a little longer before it is taken.
  • While an answer is playing, the assistant does not hear you. Wait for Listening before you speak.
  • Voice conversation and dictation exclude each other; end one before you start the other.
  • If you ask the next question while an answer is still playing, only your latest question is answered after the current answer. Earlier questions in between stay in the chat without an answer.
  • Words you speak at the very end of a recording are stored in the chat but get no spoken answer.
  • Read the sources in the chat. The assistant does not read them out.
  • The gateway's own tools keep running; web search runs only if the gateway is set to run it on every message (mode always). Connectors (MCP) and connector setup are not used in a voice conversation; ask in a typed message to use them (see My MCP Connectors).

Limits

  • The assistant does not propose new memories in a voice conversation; stored memories are still used.
  • A voice conversation is always answered by your workspace's local Myra model, even if the conversation shows a cloud model in the model picker. The picker is not changed and typed messages keep using the chosen model. The voice row shows Myra model, and the stored answer names the local model that answered. If no Myra model is enabled for this gateway, no answer is spoken. Ask your administrator to enable a Myra model.
  • Answers are spoken only in conversations you own. In a conversation shared with you through a project, your spoken turns are stored but no answer is read out.
  • An answer longer than about 64 KiB of spoken text is stored in the chat in full; speaking stops when that limit is reached.
  • Speak German, English or French. If a sentence is recognised in another language, it is skipped: it is not stored and gets no answer, and the voice session continues. Repeat the sentence in German, English or French.
  • One recording is limited to ten minutes. The assistant sees the most recent messages of the conversation, not all of a very long one.
  • Typical timing: in a test with 40 answers, the first audio started about 3.5 seconds after the end of speech (95 % within about 4.1 seconds). This depends on the setup and is not a guarantee.

When something goes wrong

  • Speaking stops or nothing is spoken; the answer is in the chat. The speech service failed or is busy for a moment. What was already spoken stays; read the rest of the answer in the chat and ask again.
  • No answer; no Myra model could answer. Spoken answers are produced by the local Myra model, and none could be used for this turn: either none is enabled for this gateway, or there was a brief fault. While the voice conversation runs the chat shows no card for the unanswered turn; after you click on End, the No answer arrived card appears under it. Ask again; if it keeps happening, ask your administrator to enable a Myra model.
  • The conversation is read-only. Someone else is driving a live conversation or you cannot write to it. Nothing is spoken. If someone takes over the live conversation after your question was stored and before the answer, your question stays in the chat and the answer ends as read-only. If someone else is already driving when your question would be stored, the session closes and the question is not stored.
  • The assistant stays silent, or stops part-way, after your question: the answer could not be produced. Your question is kept in the chat; ask again.
  • A sentence is ignored and gets no answer. It was probably recognised in a language other than German, English or French. Such sentences are skipped and not stored; the voice session keeps running. Say the sentence again in German, English or French.
  • The voice session ends with an error "Voice mode failed." This points to a fault of the speech service or the gateway, not to the language you spoke. Start the voice session again; if it keeps happening, tell your administrator.
  • There is no Voice mode button. The platform administrator has not switched voice conversation on for your workspace.