Voice

The Voice tab gives the app a voice and ears: it can read replies aloud, let you dictate instead of type, and — when a worker is on a provider that supports it — hold a live, hands-free Talk conversation. It comes in three tiers, from “free and on your machine” to “cloud, billed per minute.”

🔊 Text-to-speech (Tier 1)

Hear workers read their replies out loud, using your operating system’s built-in voices.

Auto-read assistant replies aloud (checkbox)

What it does: speaks each worker reply automatically once it finishes.
Default: off.
When to change it: turn it on for a hands-free, listen-while-you-work feel. Even with it off, you can still click the 🔊 button on any single message to hear that one on demand.

What it does: picks which of your OS’s installed voices is used.
Default: System default.
When to change it: choose a specific voice or language you prefer. (On some systems the list fills in a moment after you open the tab.)

Rate, Pitch, Volume (sliders)

What they do: Rate = speaking speed, Pitch = how high or low the voice sounds, Volume = loudness.
Defaults: middle-of-the-range, natural settings.
When to change them: speed Rate up if the default feels slow; nudge Pitch/Volume to taste.

🔊 Test voice (button)

What it does: speaks a sample sentence so you can hear your current settings immediately.

🎤 Voice input / dictation (Tier 1)

Speak your messages instead of typing them.

A status line at the top tells you what’s available on this machine: a green tick if live dictation works, an amber note if only an audio-recording fallback is available, or a red note if neither is.

Show the 🎤 mic button in the chat composer (checkbox)

What it does: adds a microphone button to the chat box. Click it to dictate; your words appear in the text box, and you press Enter when ready.
Default: off (and unavailable to switch on if your machine has no speech-recognition path at all).
When to change it: turn it on whenever you’d rather talk than type.

Use audio-recording fallback (checkbox)

What it does: when live transcription isn’t available, this records short audio chunks instead, to be transcribed by a later feature.
Default: off.
When to change it: enable it only if the status line says live dictation isn’t available and you still want to capture audio. (Transcription of these chunks arrives with the planned local-Whisper feature.)

Microphone permission (button)

What it does: triggers your OS’s microphone-permission prompt now, so the mic button works the instant you first click it in chat. It reports granted, denied, or offers to re-ask.
When to use it: click it once after enabling the mic button. If it shows denied, allow the app under your OS privacy settings (e.g. macOS → System Settings → Privacy & Security → Microphone).

🎧 Talk mode — cloud Realtime (Tier 2)

A true duplex voice conversation: you speak and hear a spoken reply, you can interrupt mid-sentence (“barge-in”), with no read-aloud delay. There’s nothing to toggle on this tab — a 🎧 Talk button simply appears in the chat composer whenever the active worker is on a provider that supports it.

OpenAI — a real-time voice link to OpenAI (gpt-4o-realtime). Sub-second latency, barge-in. The secure handshake is signed with your stored key server-side, so the key never touches the page.
Gemini — Google’s Gemini Live voice link (gemini-2.0-flash-live, voice Aoede). It replies in whatever language you speak, supports barge-in, and is far cheaper per minute than OpenAI. Reuses your Gemini key.
Anthropic — no public real-time voice yet (as of mid-2026); the button will appear here too once one exists.

🌐 Local high-quality voice — Whisper + Piper (Tier 3)

What it will do: download Whisper (speech-to-text) and Piper (text-to-speech) models from the Models tab and run them entirely on your machine — replacing the OS voices with much warmer, more natural ones, and replacing live dictation with offline transcription.
Status: planned. It’s a larger download, but stays 100% on-device.
Why later: each piece (the engine, the model catalogue entries, the streaming wiring) is real work; Tier 2 shipped first because it reuses keys you already have.

In practice

Where voice genuinely helps: dictating a long brief you’d type slowly, reviewing a draft by ear (you catch clumsy phrasing reading aloud that you skim past on screen), and hands-busy situations — a workshop, a car, a walk.

System voices cost nothing and are always available. Start there before wiring a provider-backed voice. If the OS voice is good enough for reviewing your own drafts, that’s a free capability.

Live voice bills by the minute, unlike text which bills by tokens. That’s a different mental model: a long pause in a spoken conversation still costs. Fine for a demo or a short exchange, worth watching on anything running in a public space.

Dictation quality follows the microphone. A quiet room and a decent mic matter more than the model — background noise produces transcription errors that then get treated as your instruction, which is a confusing way to get a wrong result.

Terms in this page

Text-to-speech (TTS) — turning written text into spoken audio (the app reading replies aloud).
Speech recognition / dictation (ASR) — turning your spoken words into text you can send.
System voice — one of the voices your operating system provides for reading text aloud.
Rate / Pitch / Volume — speaking speed, how high or low the voice is, and loudness.
Talk mode — a live, hands-free voice conversation with a worker.
Duplex / full-duplex — a voice mode where both sides can speak and be heard at the same time, like a phone call, rather than taking strict turns.
Barge-in — being able to interrupt the voice while it’s talking and have it stop and listen.
Realtime (cloud) — the provider technology behind Talk mode that streams audio both ways with very low delay.
gpt-4o-realtime / Gemini Live — OpenAI’s and Google’s real-time voice services, respectively.
Latency — the delay before you hear a response. “Sub-second” means it feels immediate.
Whisper — an open speech-to-text model (planned for on-device dictation).
Piper — an open text-to-speech model (planned for warmer on-device read-aloud).
On-device / local — running on your own computer, with no audio leaving it.
Microphone permission — your OS’s approval for the app to use the mic; granted once, then remembered.
Cost ledger — the running tally of what you’ve spent, shown on the Cost tab; Talk-mode minutes are added to it.