xAI adds adjustable VAD threshold to Speech to Text API

xAI updated its Speech to Text API with an adjustable voice-activity-detection threshold, giving developers finer control over how the model distinguishes speech from silence or background noise.

What's new

Per xAI's release notes: "Speech to Text now accepts a vad_threshold parameter (streaming query param and batch multipart field) to tune the voice-activity gate that skips non-speech audio. Lower values transcribe quieter or noisier speech — useful for narrowband telephony — and 0 disables the gate."

The parameter is available on both the streaming query interface and the batch multipart upload path, so it applies whether audio is processed in real time or submitted as a file. Setting it to 0 turns off the voice-activity gate entirely, meaning the model will attempt to transcribe the full audio stream rather than skipping segments it judges to be non-speech.

Context

Voice-activity detection (VAD) is a preprocessing step most speech-to-text systems use to avoid wasting compute transcribing silence or noise. A fixed threshold works fine for clean studio audio but tends to drop or mangle quieter speech in noisier conditions — a common complaint from developers building on top of telephony or call-center audio, where signal quality is inherently degraded. xAI's own note calls out narrowband telephony specifically as a use case that benefits from a lower threshold.

This is an incremental API change rather than a new model, following a string of smaller Speech to Text refinements xAI has shipped through its docs.x.ai release notes over recent weeks as it builds out the audio side of its Grok API alongside its core chat models.

Why it matters

On its own, an adjustable VAD threshold is a small, practical fix rather than a headline feature — but it matters to a specific, revenue-relevant segment of developers: those building voice agents for call centers, IVR systems, and other telephony use cases where audio quality is inherently poor. Competitors' speech APIs have shipped similar tunable-sensitivity controls, so this brings xAI's Speech to Text closer to parity on a feature enterprise voice-agent builders tend to ask for. It's the kind of unglamorous API polish that determines whether a platform is viable for production voice workloads, not just demos.