> ## Documentation Index
> Fetch the complete documentation index at: https://docs.burki.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# ElevenLabs Scribe v2

> Configure ElevenLabs Scribe v2 Realtime for low-latency speech-to-text with broad language support

<Callout type="info">
  ElevenLabs Scribe v2 Realtime is a low-latency STT provider with broad language support. It uses advanced Voice Activity Detection (VAD) for intelligent speech boundary detection; exact latency and accuracy depend on language, audio quality, and provider conditions.
</Callout>

***

## Overview

| Feature                | Details                                                |
| ---------------------- | ------------------------------------------------------ |
| **Model**              | `scribe_v2_realtime`                                   |
| **Latency**            | Vendor-reported low latency; measure in your call path |
| **Accuracy**           | Provider/model-dependent                               |
| **Languages**          | 90+                                                    |
| **VAD Support**        | ✅ Advanced                                             |
| **Word Timestamps**    | ✅ Optional                                             |
| **Language Detection** | ✅ Optional                                             |

***

## Quick Start

1. **Get API Key**: Sign up at [ElevenLabs](https://elevenlabs.io) and get your API key
2. **Add to Provider Keys**: Go to **Settings** → **Provider Keys** and add your ElevenLabs API key
3. **Configure STT**: In your assistant settings, select ElevenLabs as the STT provider

```json theme={null}
{
  "stt_settings": {
    "provider": "elevenlabs",
    "model": "scribe_v2_realtime",
    "language": "en"
  }
}
```

***

## Configuration Options

### Basic Settings

| Parameter  | Type   | Default              | Description                      |
| ---------- | ------ | -------------------- | -------------------------------- |
| `model`    | string | `scribe_v2_realtime` | The STT model to use             |
| `language` | string | `en`                 | Language code (ISO 639-1 format) |

### VAD Settings

Voice Activity Detection (VAD) controls how the STT provider detects when someone has finished speaking.

| Parameter                    | Type   | Default | Range           | Description                              |
| ---------------------------- | ------ | ------- | --------------- | ---------------------------------------- |
| `commit_strategy`            | string | `vad`   | `vad`, `manual` | How transcripts are committed            |
| `vad_threshold`              | float  | `0.4`   | 0.1 - 0.9       | VAD sensitivity (lower = more sensitive) |
| `vad_silence_threshold_secs` | float  | `1.5`   | 0.3 - 3.0       | Silence duration to commit transcript    |
| `min_speech_duration_ms`     | int    | `100`   | 50 - 2000       | Minimum speech duration to consider      |
| `min_silence_duration_ms`    | int    | `100`   | 50 - 2000       | Minimum silence duration to consider     |

### Additional Options

| Parameter                    | Type    | Default | Description                           |
| ---------------------------- | ------- | ------- | ------------------------------------- |
| `include_timestamps`         | boolean | `true`  | Include word-level timestamps         |
| `include_language_detection` | boolean | `false` | Include detected language in response |

***

## Full Configuration Example

```json theme={null}
{
  "stt_settings": {
    "provider": "elevenlabs",
    "model": "scribe_v2_realtime",
    "language": "en",
    "elevenlabs_config": {
      "commit_strategy": "vad",
      "vad_threshold": 0.4,
      "vad_silence_threshold_secs": 1.5,
      "min_speech_duration_ms": 100,
      "min_silence_duration_ms": 100,
      "include_timestamps": true,
      "include_language_detection": false
    }
  }
}
```

***

## VAD Commit Strategy

ElevenLabs Scribe v2 uses a VAD-based commit strategy by default, which automatically detects when speech has ended and commits the transcript.

<Accordion title="VAD vs Manual Commit">
  **VAD (Voice Activity Detection) - Recommended:**

  * Automatically detects speech boundaries
  * Commits transcript when silence is detected
  * Best for natural conversation flow
  * Configurable sensitivity and timing

  **Manual Commit:**

  * You control when transcripts are committed
  * Useful for specific use cases where you need precise control
  * Requires handling commit signals in your application

  <Callout type="tip">
    For phone calls, VAD commit strategy is recommended as it provides the most natural conversation experience.
  </Callout>
</Accordion>

***

## VAD Tuning Guide

<Accordion title="Adjusting VAD for Different Scenarios">
  ### Fast-Paced Conversations

  ```json theme={null}
  {
    "vad_threshold": 0.3,
    "vad_silence_threshold_secs": 0.8,
    "min_speech_duration_ms": 50,
    "min_silence_duration_ms": 50
  }
  ```

  * Lower silence threshold for quicker responses
  * More sensitive VAD detection

  ### Patient/Thoughtful Speakers

  ```json theme={null}
  {
    "vad_threshold": 0.5,
    "vad_silence_threshold_secs": 2.0,
    "min_speech_duration_ms": 150,
    "min_silence_duration_ms": 150
  }
  ```

  * Higher silence threshold to avoid cutting off
  * More patience for pauses during thinking

  ### Noisy Environments

  ```json theme={null}
  {
    "vad_threshold": 0.6,
    "vad_silence_threshold_secs": 1.5,
    "min_speech_duration_ms": 200,
    "min_silence_duration_ms": 200
  }
  ```

  * Higher VAD threshold to filter noise
  * Longer minimum durations to avoid false triggers
</Accordion>

***

## Supported Languages

ElevenLabs Scribe v2 supports 90+ languages. Use the ISO 639-1 language code:

<Accordion title="Common Language Codes">
  | Language | Code | Language   | Code |
  | -------- | ---- | ---------- | ---- |
  | English  | `en` | Spanish    | `es` |
  | French   | `fr` | German     | `de` |
  | Italian  | `it` | Portuguese | `pt` |
  | Dutch    | `nl` | Polish     | `pl` |
  | Russian  | `ru` | Japanese   | `ja` |
  | Korean   | `ko` | Chinese    | `zh` |
  | Arabic   | `ar` | Hindi      | `hi` |
  | Turkish  | `tr` | Swedish    | `sv` |

  <Callout type="info">
    For the full list of supported languages, visit the [ElevenLabs documentation](https://elevenlabs.io/docs).
  </Callout>
</Accordion>

***

## Word-Level Timestamps

When `include_timestamps` is enabled, each word in the transcript includes timing information:

```json theme={null}
{
  "text": "Hello, how can I help you today?",
  "words": [
    { "word": "Hello", "start": 0.0, "end": 0.45 },
    { "word": "how", "start": 0.52, "end": 0.68 },
    { "word": "can", "start": 0.70, "end": 0.85 },
    { "word": "I", "start": 0.87, "end": 0.92 },
    { "word": "help", "start": 0.95, "end": 1.15 },
    { "word": "you", "start": 1.18, "end": 1.32 },
    { "word": "today", "start": 1.35, "end": 1.72 }
  ]
}
```

<Callout type="tip">
  Word timestamps are useful for analytics, keyword spotting, and advanced conversation analysis.
</Callout>

***

## Best Practices

<CardGroup cols={2}>
  <Card title="Start with Defaults" icon="play">
    The default VAD settings work well for most phone call scenarios. Only adjust after testing.
  </Card>

  <Card title="Test with Real Calls" icon="phone">
    Record test calls and listen to the conversation flow. Adjust VAD settings based on actual user experience.
  </Card>

  <Card title="Match Call Style" icon="sliders">
    Fast customer service? Lower thresholds. Complex discussions? Higher thresholds and more patience.
  </Card>

  <Card title="Enable Denoising" icon="waveform">
    For noisy environments, enable audio denoising in your STT settings alongside VAD tuning.
  </Card>
</CardGroup>

***

## Troubleshooting

<Accordion title="Common Issues">
  **Transcripts cut off mid-sentence:**

  * Increase `vad_silence_threshold_secs` (try 2.0 seconds)
  * Increase `min_silence_duration_ms`

  **Long pauses before AI responds:**

  * Decrease `vad_silence_threshold_secs` (try 1.0 seconds)
  * Lower `vad_threshold` for more sensitive detection

  **Background noise triggering false transcripts:**

  * Increase `vad_threshold` (try 0.6)
  * Increase `min_speech_duration_ms`
  * Enable audio denoising

  **Wrong language detected:**

  * Set the correct language code explicitly
  * Disable `include_language_detection` if not needed

  **Connection issues:**

  * Verify your ElevenLabs API key in **Settings** → **Provider Keys**
  * Check your account has sufficient credits
</Accordion>

***

## Comparison with Other Providers

| Feature         | ElevenLabs Scribe v2        | Deepgram                    | Azure Speech              |
| --------------- | --------------------------- | --------------------------- | ------------------------- |
| Latency         | Vendor-reported low latency | Vendor-reported low latency | Provider/config dependent |
| Languages       | 90+                         | 30+                         | 100+                      |
| VAD             | Advanced                    | Basic                       | Standard                  |
| Word Timestamps | ✅                           | ✅                           | ✅                         |
| Term Boosting   | ❌                           | ✅ Keywords                  | ✅ Phrase Lists            |
| Best For        | Multi-language, VAD         | Speed, English              | Microsoft ecosystem       |

<Callout type="info">
  Choose ElevenLabs Scribe v2 when you need excellent multi-language support with advanced VAD capabilities. Choose Deepgram for the absolute lowest latency, or Azure for enterprise features and custom models.
</Callout>