# Voice & audio (STT/TTS)


# Voice & audio (STT/TTS)

The **Audio** tab (labeled **Voice** in the agent builder) controls how your agent hears and speaks. Both speech-to-text (STT) and text-to-speech (TTS) are **multi-provider** — you choose a provider, then a model and the provider-specific controls. Getting these right is what makes an agent sound natural and understand callers accurately. Set this in the agent builder at **/agent/setup**.

:::note[Pipeline mode only for STT]

The speech-to-text settings apply in **Pipeline** voice mode. In the realtime modes (Azure Realtime, Gemini Realtime) transcription is handled internally by the provider, so the STT section is ignored. Set the voice pipeline mode on the [LLM](./llm) tab.

:::

## Speech-to-text

![The Voice tab](/img/screenshots/portal/agent-audio.png)
*The Voice tab: language, supported locales, language detection, and the speech-to-text / text-to-speech provider settings.*

Speech-to-text turns the caller's audio into text the agent can act on. Choose a **provider** and **model** that fit your callers and your domain. STT providers include:

- **Deepgram** (default) — Nova-3 / Nova-2 and more.
- **Azure** — Default and Conversation models.
- **Groq** — Whisper Large v3 / Turbo.
- **Sarvam** — Saarika (Indic).

### Common controls

**Language**, **punctuation**, **word timestamps** and **interim results** apply across providers. Setting the right language is the single biggest factor in transcription accuracy.

### Provider-specific controls

Some controls only appear for the provider that supports them:

- **Deepgram** — keyword boost, smart format, filler removal, **diarization** (separate the transcript by speaker), profanity filter, alternatives and latency mode.
- **Azure** — profanity mode (masked / removed / raw).
- **Sarvam** — code mixing.

You can also add **custom vocabulary** (product names, brand terms, jargon) to improve recognition, and tune endpointing and VAD turnoff timings.

:::note[PII redaction lives on Guardrails]

Transcript **PII redaction** (which sensitive types to mask) is configured on the **Guardrails** tab, not here. See [Guardrails](./guardrails).

:::

## Text-to-speech

Text-to-speech is the voice your agent speaks with. Choose a **provider**, **model** and **voice**. TTS providers include:

- **ElevenLabs** (default) — Turbo v2.5 / v2, Multilingual v2.
- **Azure** — Neural and Standard.
- **Groq** — Orpheus.
- **Cartesia** — Sonic.
- **Google** — Gemini TTS.
- **Sarvam** — Bulbul (Indic).

### Voice

Choose the **voice** that fits your brand and your callers. You can also use a custom voice created in [Voice Lab](../voice-lab) — clone a voice from a short sample and assign it to the agent. Use the preview button to hear it.

### Common controls

**Speed**, **pitch**, **style**, **volume**, **emotion**, **emphasis**, **output format** and custom **pronunciations** apply across providers. Slightly slower speech is easier to follow on a phone call, especially for numbers, dates and confirmations.

### ElevenLabs-specific controls

**Stability**, **similarity boost**, **style exaggeration** and **speaker boost** are ElevenLabs voice controls:

- **Stability** controls how consistent the voice sounds from phrase to phrase — higher is steadier, lower allows more natural variation.
- **Similarity boost** keeps the output close to the source voice.
- **Speaker boost** enhances the resemblance to the original speaker.

:::tip[Test out loud]

Test changes out loud, not on paper. Use the **voice test** at **/agent/interface** after each adjustment — speed and pronunciation issues are obvious in seconds when you hear them.

:::

## Next steps

- [Clone and assign a voice in Voice Lab](../voice-lab)
- [Choose the voice pipeline mode and LLM](./llm)
- [Shape conversation behavior](./engine-behavior)
- [Set recording, transcription and retention](../workplace-team)