Endpointing

Endpointing controls how long the server waits after detecting the last spoken word before it marks a transcript segment as final (is_final: true). Use this parameter to balance transcription speed against sentence completeness.

Why do you need endpointing?

When speaking, it is natural to pause between words, thoughts, and sentences. A speech-to-text system needs to decide which pauses are mid-sentence breathers and which signal the end of a complete thought. This decision is called endpointing.

Without endpointing, the system would either finalize too early (splitting "Hello world" into "Hello" and "world" as separate results) or wait indefinitely for more speech. The endpointing parameter gives you control over how quickly the API finalizes sentences during streaming.

A lower value finalizes results faster but might split sentences mid-thought. A higher value waits longer and produces more complete sentences, but adds latency.

Usage

To set the endpointing value, add the endpointing parameter in the WebSocket query string. You specify the value in milliseconds:

# Faster finalization (300ms)
wss://stt-api.subq.ai/v1/listen?endpointing=300&encoding=mp3

# Slower finalization (800ms), better for speakers who pause mid-sentence
wss://stt-api.subq.ai/v1/listen?endpointing=800&encoding=mp3

To disable automatic finalization, set endpointing to false as follows:

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

When you disable endpointing, the server finalizes results only when you send a Finalize or CloseStream message. This is useful when your application needs full control over when transcription segments are committed. For example, a dictation app where the user explicitly presses a button to finish a sentence.

Choosing an endpointing value

Next steps

• Interim results
• Utterance detection
• WebSocket protocol