# xAI adds a tunable voice-activity threshold to Speech to Text

**xAI's Speech to Text API now accepts a vad_threshold parameter so developers can tune how aggressively the service filters out non-speech audio.**

## What changed

According to [xAI's release notes](https://docs.x.ai/developers/release-notes), the Speech to Text API now accepts a vad_threshold parameter, available as both a streaming query parameter and a batch multipart field, to tune the voice-activity gate that skips non-speech audio.

## What the parameter actually does

The voice-activity gate is the part of Speech to Text that decides whether a stretch of audio contains speech worth transcribing or should be treated as silence and skipped. Lowering the vad_threshold value makes that gate more permissive, so the API keeps transcribing quieter or noisier speech that a stricter gate would otherwise drop. The release notes call this useful for narrowband telephony, where phone-line audio is often quieter or more compressed than a studio recording. Setting vad_threshold to 0 disables the gate entirely, so the API stops filtering non-speech audio based on this threshold at all.

## Where to find it

xAI points developers to its Speech to Text docs for the full parameter reference. Because vad_threshold is exposed both as a streaming query parameter and as a batch multipart field, the same tuning applies whether a team is transcribing audio live as it arrives or processing recorded files in batch, without needing two separate settings to learn.

## Why a tunable threshold matters

A fixed voice-activity gate has to guess, for every customer and every audio source, where speech ends and silence or background noise begins. That single guess rarely fits every situation: a clear studio microphone and a compressed narrowband phone line do not sound alike, and a gate tuned for one can quietly drop real words spoken through the other. By exposing vad_threshold directly, xAI lets a developer match the gate to the audio source instead of accepting whatever default cutoff the API ships with, and lets them test the difference directly by comparing transcripts at different threshold values, including the fully open 0 setting.

## Key facts

- The new parameter is named vad_threshold.
- It is available as a streaming query parameter and as a batch multipart field.
- Lower values transcribe quieter or noisier speech, which xAI calls useful for narrowband telephony.
- Setting vad_threshold to 0 disables the voice-activity gate entirely.
- The full parameter reference lives in xAI's Speech to Text docs.

## What you can check or try today

If your team transcribes phone-call audio, IVR recordings, or other narrowband telephony sources, check whether your current Speech to Text integration is dropping quiet or noisy speech, then try lowering vad_threshold on a streaming or batch request to see if more of that speech gets transcribed. To confirm the gate's effect on your own audio, you can also test the extreme case by setting vad_threshold to 0 and comparing the output against your normal setting. Consult xAI's Speech to Text docs for the exact request shape before changing it in production.

## Check your understanding

- What does the vad_threshold parameter control in the Speech to Text API?
- What happens if a developer sets vad_threshold to 0?
- In which two request forms can a developer supply the vad_threshold value?

## Sources

- [xAI: Release Notes](https://docs.x.ai/developers/release-notes)

---

- Category: implement (release)
- Affects: —
- Published: 2026-07-26T07:32:10.863Z
- Original: https://docs.x.ai/developers/release-notes
- Generated by tinker news