/speech-to-text

Transcripts and Commit Strategies

Control when and how transcripts are finalized in real-time streaming.

Why Commits Matter

In real-time transcription, the model continuously refines its understanding as more audio arrives. A word that sounds like "their" might become "there" or "they're" once more context is heard. The commit mechanism lets you decide when to "lock in" the transcript.

Transcript Types

Manual Commit (Default)

You explicitly control when transcript segments finalize.

Python

async with client.speech_to_text.realtime.connect(
    model_id="scribe_v2_realtime",
) as connection:
    # Send audio
    await connection.send({
        "audio_base_64": audio_base_64,
        "sample_rate": 16000,
    })

    # Commit when ready (e.g., pause in speech, end of sentence)
    await connection.commit()

JavaScript

const connection = await client.speechToText.realtime.connect({
  modelId: "scribe_v2_realtime",
});

// Send audio
connection.send({
  audioBase64: audioBase64,
  sampleRate: 16000,
});

// Commit when ready
connection.commit();

Best Practices

• Commit every 20-30 seconds for optimal performance
• Commit during silence or logical breaks (end of sentence, speaker change)
• Auto-commit at 90 seconds if no manual commit is sent

Providing Context

Send previous text with the first audio chunk to help the model:

await connection.send({
    "audio_base_64": first_chunk,
    "sample_rate": 16000,
    "previous_text": "So as I was saying,"  # Keep under 50 characters
})

This helps with:

• Continuing conversations after reconnection
• Providing context for better accuracy
• Handling sentence fragments

Voice Activity Detection (VAD)

VAD listens for silence and automatically commits when the speaker pauses. This creates natural transcript segments that match how people actually speak - pausing between sentences and thoughts. Recommended for live microphone input.

Configuration

React (useScribe)

import { useScribe, CommitStrategy } from "@elevenlabs/react";

const scribe = useScribe({
  modelId: "scribe_v2_realtime",
  commitStrategy: CommitStrategy.VAD,
  // Optional VAD tuning:
  vadSilenceThresholdSecs: 1.5,    // Silence duration before commit
  vadThreshold: 0.4,               // Speech detection sensitivity (0-1)
  minSpeechDurationMs: 100,        // Minimum speech length required
  minSilenceDurationMs: 100,       // Minimum silence length required
});

Important: The default is CommitStrategy.MANUAL. For microphone input, always set CommitStrategy.VAD — without it, committed transcripts will never fire and the connection may drop.

JavaScript client

const connection = await client.speechToText.realtime.connect({
  modelId: "scribe_v2_realtime",
  vad: {
    silenceThresholdSecs: 1.5,    // Silence duration before commit
    threshold: 0.4,               // Speech detection sensitivity (0-1)
    minSpeechDurationMs: 100,     // Minimum speech length required
    minSilenceDurationMs: 100,    // Minimum silence length required
  },
});

Parameters

When to Use VAD

• Live microphone input
• Conversational applications
• When natural speech boundaries are preferred
• Client-side implementations

When to Use Manual Commit

• Processing audio files
• Known segment boundaries
• Maximum control over timing
• Server-side batch processing

Supported Audio Formats