Voices and mouths

To change the voice, say “slash voice asteria”. The mouth is the part that speaks: it turns each sentence of a reply into speech as soon as the brain has written it. Deepgram Aura is the default mouth. ElevenLabs is the paid alternative, and Kokoro runs on your Mac with no key.

Pick a Deepgram Aura voiceCopy Anchor Link

Any Aura-2 voice works. Setup offers these six, and aura-2-orion-en is the default:

Change it by voice, or from the shell with the full name:

Pick one by how it soundsCopy Anchor Link

Say what you want instead of a name: “slash voice a British woman”, “slash voice deep male”, “slash voice calm storytelling”. Each Aura voice carries Deepgram's own description of it: its accent, its age, words like warm, raspy or baritone, and what it suits, such as storytelling or customer service. “Woman” and “female” mean feminine, “man” and “male” mean masculine.

When one voice has every feature you said, it becomes the voice, as if you had said its name. When several do, they are listed on screen with their descriptions, and nothing changes; say one of them, or add a feature. “Slash voice” alone lists all of them, described the same way.

From the shell, claudio set voice still takes a name.

“Slash voice” sets the voice of the mouth that speaks: the Aura voice with Deepgram, elevenlabs_voice with ElevenLabs, local_voice with Kokoro. “Slash elevenlabs voice” and “slash local voice” set theirs whichever mouth speaks.

The name is not checked against Deepgram's list. A voice Deepgram does not know leaves the mouth unable to connect, and you see ! mouth lost. Deepgram's text-to-speech models page lists every voice.

Switch the mouth to ElevenLabsCopy Anchor Link

The ears are always Deepgram, so you still need a Deepgram key. Setup does not ask for the ElevenLabs key, so add it yourself:

{ "deepgram_api_key": "your_deepgram_key", "elevenlabs_api_key": "your_elevenlabs_key" }

export ELEVENLABS_API_KEY=your_elevenlabs_key

claudio set mouth elevenlabs

claudio set elevenlabs_voice pNInz6obpgDQGcFmaJgB

claudio streams from ElevenLabs' eleven_flash_v2_5 model as 24 kHz PCM (pcm_24000).

If mouth is elevenlabs but no ElevenLabs key is set, Deepgram Aura speaks instead, with no warning. The banner shows which mouth started: voice elevenlabs pNInz6ob for ElevenLabs, voice orion for Aura. “slash status” always names the Aura voice, even while ElevenLabs speaks.

To go back to Aura:

claudio set mouth deepgram

Speak with Kokoro, on your MacCopy Anchor Link

The local mouth runs Kokoro, an open voice model, on your Mac's CPU. It needs no key, and what claudio says never leaves the machine. The ears are still Deepgram, so you still need a Deepgram key.

brew install espeak-ng
uv tool install --reinstall 'claudio[local] @ git+https://github.com/ActurialCapital/claudio'

claudio set mouth local

claudio set local_voice af_heart
claudio set local_speed 1.1

claudio set local_model full

Synthesis time on an M4, per sentence:

Either way the first word comes a little later than with Aura. The banner shows voice local am_michael.

What happens when the mouth failsCopy Anchor Link

A reply is never lost. You read it as text instead:

• The speech service cannot be reached: claudio tries once more after one second. If that fails too, you see ! mouth lost: the speech service is unreachable; text only, and the reply is shown as text.
• The connection drops mid-reply: claudio opens a new connection and carries on from the sentence it was on. If that fails, you see ! mouth lost mid-reply; the rest is text only.

Each new reply tries the mouth again. Troubleshooting has the fixes.

SettingsCopy Anchor Link