> ## Documentation Index > Fetch the complete documentation index at: https://www.bolna.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # ElevenLabs Voice Synthesis for Bolna Agents > Configure ElevenLabs text-to-speech in Bolna voice agents, including Turbo v2.5, Flash v2.5, and Eleven v3 — models, voices, streaming, and cloning. [ElevenLabs](https://elevenlabs.io/) is Bolna's default voice synthesizer. The `eleven_turbo_v2_5` model delivers natural-sounding English speech with the lowest latency of any ElevenLabs model. *** ## Quick config ```json theme={"system"} "synthesizer": { "provider": "elevenlabs", "provider_config": { "voice": "Nila", "voice_id": "V9LCAAi4tTlqe9JadbCo", "model": "eleven_turbo_v2_5" }, "stream": true, "buffer_size": 250, "audio_format": "wav" } ``` To use your own ElevenLabs account (for voice cloning or custom voices), connect it at [platform.bolna.ai/auth/elevenlabs](https://platform.bolna.ai/auth/elevenlabs). *** ## Supported models | Model | Best for | | -------------------------- | -------------------------------------- | | `eleven_turbo_v2_5` | Low-latency English (recommended) | | `eleven_flash_v2_5` | Fastest option; slightly lower quality | | `eleven_v3_conversational` | Most expressive; 74 languages | `eleven_turbo_v2_5` is the standard choice for production agents. Use `eleven_flash_v2_5` when you need the absolute lowest time-to-first-audio. `eleven_v3_conversational` is ElevenLabs' most expressive model and supports 74 languages. It ignores `speed`, `style` and `similarity_boost`, and accepts only three `temperature` values (see below). *** ## Key settings | Setting | Type | Default | Description | | -------------- | ------- | ------- | ----------------------------------------------------------------------------------------- | | `voice` | string | — | Display name of the voice | | `voice_id` | string | — | ElevenLabs voice ID (stable; use this, not the name) | | `model` | string | — | ElevenLabs model (`eleven_turbo_v2_5`, `eleven_flash_v2_5` or `eleven_v3_conversational`) | | `stream` | bool | `true` | Enable streaming; always `true` for real-time calls | | `buffer_size` | integer | `250` | Characters to buffer before sending first audio chunk | | `audio_format` | string | `wav` | Output format — match your telephony provider | ### Eleven v3 settings `eleven_v3_conversational` takes a narrower set of voice settings than the Turbo and Flash models: | Setting | Behaviour on v3 | | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------- | | `temperature` | Maps to stability, which v3 treats as three presets: `0.0` creative, `0.5` natural, `1.0` robust. Other values are rounded to the nearest preset. | | `speed` | Ignored | | `style` | Ignored | | `similarity_boost` | Ignored | ### buffer\_size guidance `buffer_size` controls the trade-off between first-word latency and audio smoothness: * **100–150** — fastest first word; can sound choppy if ElevenLabs response is slow * **250** — balanced default; works well for most agents * **400+** — smoothest speech; higher latency to first word *** ## Choosing a voice ElevenLabs voices are identified by a `voice_id`. The `voice` name field is for reference only — Bolna uses `voice_id` to select the voice. Browse voices in the [ElevenLabs Voice Library](https://elevenlabs.io/voice-library). Copy the voice ID from the URL or the voice details panel. **Example voices:** | Voice | voice\_id | Accent | | ------ | ---------------------- | ---------------- | | Nila | `V9LCAAi4tTlqe9JadbCo` | Neutral American | | Rachel | `21m00Tcm4TlvDq8ikWAM` | Neutral American | | Adam | `pNInz6obpgDQGcFmaJgB` | Neutral American | For a custom or cloned voice, use the voice\_id from your ElevenLabs account. See [Clone Voices](/docs/clone-voices) and [Import Voices](/docs/import-voices). *** ## audio\_format by telephony provider | Telephony | audio\_format | | ---------- | ---------------------- | | Plivo | `wav` | | Exotel | `wav` | | Twilio | `mp3` or `wav` | | Custom SIP | match codec negotiated | *** ## FAQ Use `eleven_turbo_v2_5` for most production agents. It has the better quality/latency balance. Use `eleven_flash_v2_5` only if your primary constraint is time-to-first-audio and you've measured that Turbo is too slow. Use `eleven_v3_conversational` when expressiveness is the priority; it is the most expressive of the three, and Turbo remains the lower-latency option. `eleven_v3_conversational` ignores `speed`, `style` and `similarity_boost`; the model does not act on them. Only `temperature` (stability) has an effect, and it accepts three values: `0.0`, `0.5` and `1.0`. Anything else is rounded to the nearest of those. Use Turbo or Flash if you need speed control. In the ElevenLabs dashboard, open any voice and look at the URL — the ID is the path component after `/voice/`. Alternatively, list voices via the ElevenLabs API. Always store the `voice_id`, not just the name; names can change. Yes. Clone your voice in ElevenLabs, connect your ElevenLabs account to Bolna, then set `voice_id` to your cloned voice's ID. See [Clone Voices](/docs/clone-voices). Increase `buffer_size` to 350–400. Choppy audio usually means the first audio chunk starts before ElevenLabs has generated enough audio to stream smoothly. *** ## Related * [Audio Tab](/docs/agent-setup/audio-tab) — configure synthesizer in the dashboard * [Clone Voices](/docs/clone-voices) — create a custom voice * [Import Voices](/docs/import-voices) — use existing ElevenLabs voices * [Cartesia](/docs/integrations/voice/cartesia) — alternative low-latency synthesizer * [Latency](/docs/concepts/latency) — how synthesis affects response time