Skip to main content
Noise Cancellation is part of the STT Performance Layer in PRIVATE BETA. The behavior described here is being rolled out gradually. Contact us for access.
STT accuracy degrades fast in noisy audio. Most voice AI calls happen in noisy environments: call centers, mobile phones on the street, motorbikes, background TV, shared rooms. Every point of word error rate lost to background noise compounds through the whole Agent experience. That compounding is the real argument. In a voice agent, STT output feeds the LLM. A garbled transcript produces a wrong intent, which produces a wrong response, which produces a user repeating themselves or dropping off. The cost of bad audio does not show up as a transcription metric, it shows up as agent failure rate and containment rate. Cleaning the audio at the front of the chain is key for the success of the call.

​
How it works

SLNG Noise cancellation works across all the STT models that are available in the SLNG gateway. To activate it you need to pass a parameter through the header in the init message of your websocket connection. Request level : users activate or deactivate it per request with a single header on the WebSocket upgrade:
No header = noise cancellation skipped, audio passes through unchanged.

​
Models applied

The model being used for the Noise cancellation pipeline is Aicoustics, at SLNG we did an exhaustive research to find the best one and the Quail model came on top. You can find the information of the whitepaper here: SLNG Noise Cancellation Whitepaper

​
Failure Behavior

Fail-open by default. If anything goes wrong, the original audio is forwarded to the STT provider unchanged. The client never gets an error because of noise cancellation.