Voice calls

Talk to your agents out loud instead of typing. In a one-to-one chat you speak and the agent answers in a synthetic voice; in a group chat each member gets its own voice and speaks in turn.

Voice is an optional add-on, and it is not released yet. Without it, chats show no voice controls; only the Voice Calls card in Settings appears, showing that the add-on is not installed. The rest of this page describes how voice calls work once the add-on is available.

What you need

The Verzeta-Voice add-on, a separate download.

It is a separate program because it is large (speech models and their runtimes) and licensed differently (GPL-3.0-or-later). It runs on your machine and talks to Verzeta Studio locally. Nothing about your voice or your conversations leaves your computer.

You also need a microphone, and speakers or headphones.

Installing

Build or download verzeta-voice (see that project's README).
Install it, or copy it to one of these folders. Verzeta Studio scans the usual places automatically: a system install (/usr/bin, /usr/local/bin, Program Files), /opt/verzeta-voice, ~/.local/bin, ~/Applications, your Downloads folder, and the per-user folder shown in Settings → Voice Calls. A Linux AppImage is detected by name, so Verzeta-Voice-1.0.0-x86_64.AppImage works as downloaded with no renaming. Any folder on your PATH also works.
Press Detect. The line turns green and shows where it was found. You do not need to restart Verzeta Studio.

If it is installed somewhere unusual, press Locate… and pick the program in the file dialog. That button only appears while nothing has been detected, and Clear custom path returns to the normal search.

Speech models

The add-on ships without models, so you choose the quality and size you want. Download them in Manage Voice Calls → Downloads (see below), or follow the download commands in the add-on's README and put them in the add-on's models folder (~/.local/share/verzeta-voice/models on Linux). You need two things:

one speech-to-text model (a whisper ggml-*.bin file), and
at least one voice for speech output (a piper .onnx file with its matching .onnx.json).

Add more voices to give each agent a different one.

Turning it on

The Voice Calls section in Settings shows whether the add-on is installed and running, and Manage Voice Calls opens everything else: a switch to start and stop the service, Detect and Locate for finding the program, the start-with-application option, the default voice, downloads, and the push-to-talk key.

If the add-on is a different version than Verzeta Studio expects, the status line says so and names both versions. Update the add-on.

Making a call

A call button appears in the chat header once the service is running. Click it and the call window opens over the chat, showing a tile for you and one for each agent.

Hold to talk. Press and hold the talk button while you speak, then release. Holding it also stops whatever an agent is saying, so this is how you interrupt: press, speak, release. The agent stops talking and answers what you just said instead of finishing first.
Stop speaking cuts an agent off without you having to say anything, for when you can already tell the answer is going the wrong way. Only the audio stops; the reply stays in the chat where you can read it.
Who you are talking to. In a group chat, the dropdown beside the talk button chooses one agent or everyone. Verzeta adds the mention for you, so you never have to say "at" out loud. The choice sticks until you change it.
Minimize collapses the call to a small bar so you can read and use the chat while the call carries on. The bar keeps the talk button, stop speaking, and hang up. Click restore to bring the full view back.
End call closes the call. The conversation stays; everything said is in the chat as ordinary messages.

Only one call can run at a time.

There is no mute button. The microphone is open only while you hold the talk button, so letting go already means nothing is being heard.

Push to talk, and why

The microphone is open only while you hold the talk button. That keeps the agents from hearing themselves through your speakers and answering their own voices. Always-open listening needs echo cancellation and is planned for a later release.

How group calls take turns

In a group call each agent finishes speaking before the next one starts thinking. This is deliberate: it keeps what you hear in the same order as the conversation, so when you interrupt, the reply you get is to what you said. It does mean a small pause between speakers while the next agent composes its answer.

A one-to-one call does not wait like this. There is no next speaker to order against, so you can talk or type at any point, including while the agent is still reading its answer out.

Choosing voices

Voice choices live with the rest of an agent's settings:

The default voice is in Settings → Voice Calls → Manage Voice Calls. Any agent without a voice of its own uses it.
In a group chat, open Chat Settings → Group Members, press the configure button on a member, and pick its Voice next to its model override. The choice applies to that member in that chat.
In a one-to-one chat, Chat Settings has a Voice section right under Model.

The voice controls appear only while the voice service is running, and a change during a call is heard on that agent's next spoken reply.

Downloading voices and speech models

Manage Voice Calls has a Downloads panel with ready-to-use voices (male and female, US and UK English) and better speech-recognition models. Each row shows the size and the licence; every download is checked against a known fingerprint before it is installed, and the voice service restarts by itself so the new voice appears right away.

A larger speech-recognition model understands you noticeably better than the starter one, and the best installed model is used automatically on the next call.

The push-to-talk key

The on-screen talk button always works. If you prefer a key, set one in Manage Voice Calls → Push to talk: click the field and press the key combination. Hold it during a call to talk, exactly like holding the button. The key is active only while a call is running and never fires while you are typing in a text field.

What gets spoken

Only what an agent says, and only the parts worth hearing.

Code blocks are read as "Code omitted." A table is read as a one-line summary of its columns and row count, for example "A table of Feature and Price with 3 rows is on screen." Read out in full they would be a stream of punctuation. Both stay fully visible in the chat. Tool calls, thinking, and system notices are never spoken, and a turn that only ran a tool passes silently.

Troubleshooting

"Verzeta-Voice is not detected" after installing. Check that the program is in one of the folders named in Settings, and that it is executable. On Linux: chmod +x verzeta-voice. Press Detect again.

"Verzeta-Voice speaks protocol vX but this app needs vY". The add-on and Verzeta Studio must agree on the protocol version. Update the add-on to the release that matches your Verzeta Studio version.

"Speech models are missing" when you start a call. The add-on found no whisper model or no voice. Download them in Manage Voice Calls → Downloads, or see "Speech models" above; the message names the exact folder it looked in.

"Voice service start failed repeatedly". Verzeta Studio stops retrying after three failed starts. Fix the cause shown in the log (see "Collecting a log" below), then press Detect to try again.

"Start the voice service before calling." Turn the service on in Settings → Voice Calls → Manage Voice Calls, or turn on Start the voice service with the application.

"End the current call first." Only one call can run at a time. End the call in the other chat first.

Nothing is heard. Check your system output device and volume. In the call window, an agent's tile lights up while it is speaking; if tiles light up but you hear nothing, the problem is the output device.

Nothing is transcribed. Check the microphone in your system settings. Hold the talk button for at least half a second. Presses shorter than about a third of a second are ignored on purpose, so an accidental tap sends nothing.

Sound comes from one speaker only. Update the add-on. A single-channel stream plays on the left side only on some systems.

Every agent sounds the same. Only one voice is installed. See "Choosing voices" above.

The agents interrupt themselves. Use headphones. Speaker audio reaching the microphone is what the push-to-talk design avoids, but a very loud room can still confuse recognition.

Collecting a log for a bug report. Start Verzeta Studio with detailed logging (see Troubleshooting) and add the voice area:

QT_LOGGING_RULES="verzeta.voice*=true" ./verzeta-studio

The add-on writes its own diagnostics to its error output, which is captured in the same log.

What's next

Troubleshooting: general troubleshooting.
Multi-agent teams: building the team you will be talking to.