Voice conversation
Voice conversation lets people talk to Qry from the chat prompt and hear a short spoken answer. It is off by default: switching it on sends people's audio to the voice provider you choose.
It is not the same as dictation, which only turns speech into the text of a prompt.
How it works
- A real-time voice model holds the spoken conversation. It has one thing it can do: pass the person's question to Qry's chat.
- Each question runs as an ordinary chat turn of the conversation on screen — the model the person selected, the conversation's history, the workspace and datasource, the skills, and that person's permissions (Qry's grants, ABAC, and on a datasource that runs queries as each user, their own identity). There is no second way into the data to secure.
- The full answer is written in the conversation; the voice sums it up in a sentence or two.
- The voice describes charts in words and never asks people to look at the screen, so it works for people who cannot see it. Asking for a chart or a table out loud draws it in the conversation.
- People end a conversation with End, Esc, or by saying goodbye ("bye", "that's all, thanks"; a bare "thanks" does not end it). The voice only hangs up right after the person has spoken, so an answer that quotes "end the conversation" cannot close it.
- While the chat works, the voice says every so often what Qry is doing (querying the data, preparing the table), so a long analysis is not a long silence. Gemini Live on AI Studio cannot do this and stays silent until the answer.
- The browser only talks to Qry. Qry holds the connection to the provider, so no provider credential reaches the browser.
Where to configure
Admin › System configuration › Platform › Voice conversation.
| Setting | Meaning |
|---|---|
| Enable voice conversation | Shows the Talk button in the chat for everyone (the waveform in an empty prompt; a labelled Talk on phones). |
| Voice provider | Azure OpenAI Realtime (recommended), Gemini Live on Vertex AI, or Gemini Live on AI Studio. |
| Azure realtime deployment | The name of the realtime model's deployment in the Azure AI Foundry resource set in AI Models (for example gpt-realtime-2.1). |
| Azure voice | The voice it speaks with (default marin). |
| Gemini Live model / Vertex AI location | For the Gemini providers. |
| Max conversation length | Minutes after which a conversation closes (default 15); people can start another. |
| Answer wait | Seconds the voice waits for the chat's answer before saying it could not finish (default 180). |
| Conversations per person | How many voice conversations one person may have open at once, in different tabs or windows (default 2; 0 = no limit). |
| Conversations in total | How many may be open at once across the installation (default 0 = no limit). Set it to stay within your provider's concurrent-session quota. |
Voice conversation is hidden, like every other AI feature, when Application › Hide AI features is on.
Providers
Azure OpenAI Realtime (recommended)
- Uses the Azure AI Foundry resource from System configuration › AI Models — its key, or the pod's Entra ID identity when no key is set.
- Deploy
gpt-realtime-2.1as DataZoneStandard in an EU region to keep audio processing in the EU data zone.
Gemini Live
- Vertex AI uses the installation's Google Cloud credentials. The current Live models run in
us-central1only — the audio leaves the EU. - AI Studio uses the Gemini API key from AI Models.
Before switching it on
- Data residency: the audio and the answers' text go to the provider. Choose the provider and region your organisation's policy allows.
- Cost: the provider charges for the audio in and out. Each conversation records the tokens the provider reports in Usage, under the voice model's name, and counts against budgets like the chat does — once the model has a price. The voice model is not in Admin › Models; price it with
qry pricing apply --apply, andqry doctorreports it when it has none. A realtime model charges audio far above text —gpt-realtime-2.1is 32 / 64 USD per million tokens in and out for audio, 4 / 24 for text — so Usage shows each session twice: its audio under the model's name and its text (the voice's instructions and the answers it sums up, most of the input) undergpt-realtime-2.1:text.qry pricing applysets both. A question answered by voice measured about 0.02 USD on top of its chat turn. Until then it is recorded at zero cost. The chat turns cost what typed questions cost. - Azure deployment name: usage is recorded under the deployment's name, so name it after the model (
gpt-realtime-2.1); a deployment named otherwise has no published price to apply, and needs a price for<name>and for<name>:text. - Headphones work best in open offices: without echo cancellation the answer coming out of the speakers can interrupt itself.
Watching it
Admin › System › Platform health lists every conversation that failed — the provider refused the connection or stopped answering — under Voice conversation, with the provider's reason. A conversation the person ended — with the button or by saying goodbye — or that reached its time limit, is not a failure.
Common issues
People see "You already have voice conversations open" or "Too many voice conversations are open". They reached Conversations per person or Conversations in total. Each open conversation is a session the provider charges for. A conversation closed abruptly (a browser crash) stops counting after Max conversation length plus two minutes.
People see "Voice conversation is not set up yet". It is enabled but the provider is missing something: for Azure, the Foundry resource in AI Models or the deployment name; for Vertex, a GCP project.
"The voice service did not answer." The provider refused the connection: check the deployment name, the key or the Entra role on the Foundry resource, and that the deployment exists in that resource.
Answers stop with "Qry did not finish in time". The chat took longer than Answer wait. The answer still appears on screen; raise the wait if long analyses are normal.
See also
- Talking to Qry — the user's side.
- Speech-to-text providers — dictation.
- LLM providers — the Azure AI Foundry settings voice conversation shares.