ElevenLabs
Aiur uses ElevenLabs for Stream Deck and Dashboard voice input, and for spoken Dashboard replies.
What voice does
| Capability | Where | What happens |
|---|---|---|
| Speech to text | Stream Deck Mic key and Dashboard agent composer | Dictation is transcribed and delivered through the same agent-message path as typed Dashboard text. |
| Text to speech | Dashboard interactive conversation | The agent's reply is streamed back to the browser as speech. |
Aiur holds the credential and makes every ElevenLabs call; neither the Stream Deck sidecar nor the browser ever receives it.
API key permissions
| Permission | Needed for | Why |
|---|---|---|
Speech to Text | Dictation | Transcribes operator speech. |
User | Units meter | Reads account subscription data. |
Text to Speech | Dashboard spoken replies | Renders the agent reply as audio; also needs elevenlabs.voice_id. |
Use a restricted key with those permissions. A key that can transcribe but cannot read User data makes voice input work while the meter reports an authorization failure.
Configure the key
| Location | Value |
|---|---|
| Private environment | ELEVENLABS_API_KEY=<restricted key> |
.aiur/config | elevenlabs.api_key: $ELEVENLABS_API_KEY |
| Language | elevenlabs.language_code: eng by default. |
| Voice | elevenlabs.voice_id: null by default; set a stock or owned voice to enable Dashboard spoken replies. |
The key is optional. Without it, microphone discovery and level meters remain available, but no audio leaves the machine and transcription stays disabled.
What the Units meter measures
| Figure | Meaning |
|---|---|
| Credit quota | Account character_count against character_limit from GET /v1/user/subscription, shown as percentage used. |
| Character pool | Text-to-speech characters reported by the ElevenLabs subscription. |
| Speech-to-text cost | Not represented; ElevenLabs bills transcription per audio-minute. |
| Zero character limit | Empty track, because there is no denominator for a percentage. |
The Units meter shows only the credit quota bar; the next-invoice amount is not rendered on the strip.
The meter can remain unchanged after heavy dictation because it reads the text-to-speech character pool, not audio-minute usage.
Privacy and secret handling
| State | Data path |
|---|---|
| Dictation held open | Microphone audio goes to ElevenLabs and the returned text goes to the selected agent. |
| Dictation released | Capture stops; there is no always-on listener or wake word. |
| Key absent | No ElevenLabs connection opens and no audio leaves the machine. |
| Agent process | ELEVENLABS_API_KEY is scrubbed from coding-agent environments and never logged. |
See Stream Deck voice input for operator controls and Configuration for field defaults.