Utterance detection

Utterance detection identifies when a speaker has finished a turn by monitoring for a period of silence. When the silence period exceeds a threshold you specify, the server sends an UtteranceEnd message.

Why do you need utterance detection?

An utterance is a continuous segment of speech, that contains everything a person says before they stop talking. It might be a single sentence, several sentences, or even a few words. The key distinction is the silence that follows. For instance, when a speaker pauses long enough, that silence marks the boundary between one utterance and the next.

In a conversation, utterances roughly correspond to "turns." For example, in a voice assistant interaction, the user's question is one utterance and the assistant's response is another. You can use utterance detection to segment conversations, trigger processing between turns, or build turn-based interaction flows.

Configure the silence threshold

To enable utterance detection, add the utterance_end_ms parameter in the WebSocket query string. You also specify the silence duration that should trigger an utterance end in milliseconds:

wss://stt-api.subq.ai/v1/listen?utterance_end_ms=1000&encoding=mp3

This example fires an UtteranceEnd event after 1 second of silence. You choose a value based on your use case. A shorter threshold such as 500 ms responds quickly but might trigger during natural mid-thought pauses. On the other hand, a longer threshold such as 1500 ms is more reliable for detecting true turn endings but adds delay.

UtteranceEnd message

When the silence threshold is reached, the server sends a message like this:

{
  "type": "UtteranceEnd",
  "channel": [0],
  "last_word_end": 2.5
}

Utterance detection vs. endpointing

Both utterance detection and endpointing respond to silence, but they serve different purposes. Endpointing finalizes individual transcript segments or sentences within a turn. Utterance detection tells you when the entire turn is over. You can use both together such that endpointing finalizes sentences within a turn, and utterance detection tells you when the entire turn is over.

Related content

• Endpointing to fine-tune sentence finalization timing
• VAD events to detect when speech starts
• WebSocket protocol