Telnyx now supports three noise suppression providers across telephony, mobile, and WebRTC audio streams: DeepFilterNet, Krisp VIVA, and AI-coustics quail. Each model targets a different point in the audio pipeline, so teams can choose the option that best fits their agent architecture.
This builds on the full Telnyx communications stack, which already covers SIP trunking, programmable voice, WebRTC and Voice AI Assistants. Noise suppression is now a first-class layer across all of those surfaces, configurable through the same portal and APIs you already use.
Cleaner audio upstream means higher STT accuracy, fewer false interruptions, and more reliable turn-taking for AI agents, directly improving transcription, intent detection, and conversational flow in production.
DeepFilterNet: Open-source, full-band (48 kHz) noise suppression. Removes stationary and non-stationary background noise for telephony and WebRTC streams in a lightweight two-stage pipeline.
Krisp Viva Tel Lite: Designed for telephony use cases (up to 16kHz). Isolates the primary speaker's voice.
Krisp Viva Pro: Full voice isolation and background noise cancellation for mobile and desktop browsers (WebRTC 16-32 kHz). Removes background speech and noises, isolating only the primary speaker's voice. The model requires that the user speak close to the microphone.
Krisp Viva SS: Optimized for devices where the microphone is positioned away from the primary speaker. Smart-speaker and far-field audio variant for mobile and desktop browser (WebRTC 16-32 kHz).
AI-coustics quail: STT-optimized speech enhancement. Preserves machine-relevant phonetic structure rather than optimizing for human perception, delivering up to 43% word error rate reduction across leading transcription providers.
You can set the desired model via the noise_suppression parameter in media and call configuration requests or configure noise suppression for AI agents via the portal.
Listen to the impact of our network-native noise suppression models. Play the reference file to hear the raw, unprocessed background noise, then select a model below to hear how it isolates the voice.
Designed for machine hearing
AI agents process audio differently than humans do. They need clean phonetic structure, not natural-sounding audio. These engines are tuned for that. Processed audio may sound different to a human listener, but that is by design. The goal is better transcription accuracy, more reliable turn-taking, and fewer false triggers downstream.
To enable via the Voice API, call suppression_start on an active call. Set direction to inbound, outbound, or both, and pass noise_suppression_engine with your chosen engine name.
Learn more in our developer documentation. Need help building? Contact our team.
