elevenlabs-voice

## Overview

ElevenLabs provides text-to-speech voices for both **in-app TTS** and **phone calls**. Change the voice with the **`voice_config_update`** tool — it writes the voice to **whichever TTS provider is currently active** and pushes to the macOS app via SSE in one call:

```
voice_config_update setting="tts_voice_id" value="<voice-id>"
```

> **The voice lives under the _active_ provider, not always ElevenLabs.** The config key depends on `services.tts.provider`: `elevenlabs` → `services.tts.providers.elevenlabs.voiceId`, `vellum` (managed) → `services.tts.providers.vellum.model`, `deepgram` → `services.tts.providers.deepgram.model`. The `voice_config_update` tool (and the `assistant tts voice <id>` CLI command) handle this routing for you. **Do NOT `assistant config set services.tts.providers.elevenlabs.voiceId ...` blindly** — on a managed (`vellum`) assistant that field is ignored, so the write "succeeds" but the voice never changes. See [Setting the voice](#setting-the-voice) for the CLI fallback.
>
> **The tables below apply when the active provider is `elevenlabs` (BYO key) or managed `vellum`.** On managed assistants they are the _only_ ElevenLabs voices — the platform bills per rate-carded model and rejects anything else at synthesis (the write succeeds but the voice fails on the next turn), so don't offer library or cloned voices unless the `elevenlabs` provider is active with its own API key. With a BYO key ([setup below](#elevenlabs-api-key-setup)) and `elevenlabs` active, any voice id works. Other BYO TTS providers (`deepgram`, `xai`, `fish-audio`, …) use their own voice/model identifiers — never write an ElevenLabs voice id to them; see [Getting to an ElevenLabs voice from another provider](#getting-to-an-elevenlabs-voice-from-another-provider).

## Getting to an ElevenLabs voice from another provider

Check the active provider first: `assistant config get services.tts.provider`.

- **Already on managed `vellum`?** No provider change needed. The managed platform supports **both ElevenLabs and Deepgram voices** — `services.tts.providers.vellum.model` accepts either an ElevenLabs voice id or an Aura model id, so switching between them is just another `voice_config_update` call.
- **On a BYO provider (e.g. `deepgram`) and the user wants an ElevenLabs voice?** Two options — ask which they prefer. Switch with the `voice_config_update` tool, not raw `assistant config set` — the tool validates the switch (e.g. rejects `vellum` when no platform connection exists, which a raw config write would leave silently broken):
  1. **Switch to managed `vellum`** (`voice_config_update setting="tts_provider" value="vellum"`) — no ElevenLabs key needed; requires a platform connection and bills managed credits. Bonus: they keep access to both the ElevenLabs and Deepgram catalogs.
  2. **Switch to BYO `elevenlabs`** (`voice_config_update setting="tts_provider" value="elevenlabs"`) — requires an ElevenLabs API key ([setup below](#elevenlabs-api-key-setup)); usage bills their ElevenLabs account directly.

After either switch, set the voice with `voice_config_update` as usual.

## Choose a Voice

Pick a voice that matches your identity and the user's preferences. Offer to show the full list if they want to choose themselves.

### Female voices

| Voice   | Style                     | Voice ID               |
| ------- | ------------------------- | ---------------------- |
| Sarah   | Soft, young, approachable | `EXAVITQu4vr4xnSDxMaL` |
| Alice   | Confident, British        | `Xb7hH8MSUJpSbSDYk0k2` |
| Matilda | Warm, friendly, young     | `XrExE9yKIg1WjnnlVkGX` |
| Lily    | Warm, British             | `pFZP5JQG7iQjIQuC4Bku` |

### Male voices

| Voice   | Style                           | Voice ID               |
| ------- | ------------------------------- | ---------------------- |
| Adam    | Deep, middle-aged, professional | `pNInz6obpgDQGcFmaJgB` |
| Bill    | Trustworthy, American           | `pqHfZKP75CvOlQylNhV4` |
| George  | Warm, British, distinguished    | `JBFqnCBsd6RMkjVDRZzb` |
| Daniel  | Authoritative, British          | `onwK4e9ZLuTAKqWW03F9` |
| Charlie | Casual, Australian              | `IKne3meq5aSn9XLyUdCD` |
| Liam    | Young, articulate               | `TX3LPaxmHKxFdv7VOQHJ` |

> These are ElevenLabs' **current premade** voices. Do not use retired legacy ids (Antoni, Josh, Arnold, Rachel, Charlotte, Amelia, …): ElevenLabs silently remaps them to _different_ voices — synthesis succeeds but speaks as the wrong voice.

### Setting the voice

**Preferred — the tool.** It writes to the active provider's voice field **and** pushes to the macOS app via SSE (`ttsVoiceId`) in one call:

```
voice_config_update setting="tts_voice_id" value="<selected-voice-id>"
```

**CLI fallback (only if the `voice_config_update` tool is unavailable).** Use `assistant tts voice`, which routes to the active provider's config key for you — do **not** hand-write `assistant config set services.tts.providers.elevenlabs.voiceId ...`:

```bash
assistant tts voice "<selected-voice-id>"
```

Setting `services.tts.providers.elevenlabs.voiceId` directly while the active provider is `vellum` (or any non-elevenlabs provider) is the #1 cause of "I changed the voice but it didn't change" — that field is ignored by the active provider, so the write reports success but nothing changes. If you must use `config set`, first check `assistant config get services.tts.provider` and write the matching key (`vellum` → `services.tts.providers.vellum.model`, `deepgram` → `services.tts.providers.deepgram.model`).

Verify it worked by reading back the key for the **active** provider, e.g. for a managed assistant:

```bash
assistant config get services.tts.providers.vellum.model
```

The change hot-applies to the next voice turn (live voice and phone read the config fresh each turn).

Tell the user what voice you chose and why, but also offer to show all available voices so they can choose for themselves.

## ElevenLabs AP