Skip to content

BERD TASK · 0015 · DO THE WORK

Hold a voice conversation

Talk to the agent instead of typing at it. Voice Conversation is a two-way call with the agent in the current chat: you speak, it works, it speaks the reply while the text is still streaming. It needs a Goose session, and the first decision — Local or Cloud — decides where your voice goes. The outcome is a call you can start, leave running while you do something else, and end cleanly.

Steps at a glance

  1. Start from a Goose chat
  2. Choose Local or Cloud in the voice settings
  3. Start the call from the composer
  4. Keep working while the call runs
  5. Want text, not a call? Dictation is not in the public build

1. Start from a Goose chat

Voice Conversation is one of the two Berd features that depend on the harness: the composer's own message is Voice Conversation requires Goose. A session on Claude, Codex, Copilot, Amp, Cursor or Pi will not offer it. If you are in the wrong harness, start a new chat rather than looking for a switch — see Choose a harness for a session. The strings also carry an experiment toggle for voice conversations; if the control is missing from a Goose chat, the experiments section of Settings is the place to look before assuming a defect.

2. Choose Local or Cloud in the voice settings

The voice section of Settings has two layers. The voice mode is who you talk to: Talk to your coding agent (your speech becomes a message, the reply is read aloud; the default) or Talk through a voice assistant (an OpenAI Realtime assistant that consults the agent when it needs your project; an experiment, and it needs an OpenAI API key).

Under that, the speech backends. Since v0.6.4 they carry Local and Cloud labels: Apple speech recognition and Apple (Siri) voices run on the Mac and are the recommended default; OpenAI speech-to-text and text-to-speech send audio to OpenAI and use a voice API key that is separate from the provider credentials Goose uses. A local model may need downloading before the first call, and the settings say so plainly when something is not ready. Reset to defaults restores the recommended set-up and keeps your keys and downloaded models.

[SCREENSHOT: T15-01 — voice settings, Local and Cloud]

Also here: the interruption mode. Automatic allows interruptions on most audio devices but pauses listening on built-in Mac speakers; Allow interruptions keeps listening everywhere at the cost of the agent occasionally hearing itself; Prevent feedback never listens while it speaks, so you cannot interrupt.

3. Start the call from the composer

Start the voice conversation from the composer toolbar. Berd asks for microphone access the first time; if macOS has it turned off, the fix is in System Settings, not in Berd. The status shows listening, speech detected, agent working, agent speaking. Since v0.6.3 the reply is spoken as it streams, and v0.6.4 made it speak whole paragraphs rather than stray opening words. Each spoken reply is marked in the transcript as spoken, interrupted, not spoken or failed.

4. Keep working while the call runs

The call does not end when you leave the chat. Switch to another session or another app and floating controls follow you: mute, hang up, and a way back to the active chat.

[SCREENSHOT: T15-02 — the floating call controls]

On supported macOS devices, the mute control on AirPods and Beats toggles Berd's microphone. Archiving a chat mid-call is refused until the voice session has stopped. A voice update that could not be delivered stops the call and tells you; start it again to continue.

5. Want text, not a call? Dictation is not in the public build

Dictation — a microphone button that transcribes what you say into the message box, with "submit" sending it — exists in the source but the v0.6.4 public release compiles it off (VITE_VOICE_DICTATION=0 in the release workflow), so the button is not in the composer. Morgan, Berd's design lead, on Block's On The Block (September 2026), says "always dictate, never type" when describing what you want built; she runs Block's internal distribution, which has it. In the public build, Voice Conversation is the way to say a long brief aloud: the agent hears it as a message and works from it, and the transcript stays in the chat. In distributions that enable dictation, it needs an OpenAI key in the environment (OPENAI_API_KEY) and Berd says so on the button when the key is missing.

Latency, how well interruptions land, and how the call feels over a long session are not described in any source; they are checked at A4.

Watch out

Cloud backends send your speech to OpenAI. On an engagement with client material, choose Local before the first call, and treat what you say aloud with the same care as what you paste. See What stays on your laptop.

Last checked: Berd v0.6.4, 2026-09-11 — read from source and release notes, not yet tested in the app.

Learn more

Reference

Moment
apply
Takes
untimed
Steps
5
Verified
Berd v0.6.4 · 2026-09-11 · source-read
NEVER BESTUCK AGAINCLICK ME

Guide built with Berd, Claude and Codex · verified against Berd v0.6.4, 2026-09-11 · read from source, not yet tested in the app · Buzz Guide · About · not affiliated with Block