Inside CrystalSound: how we clean up noisy calls in real time
Introduction
Every contact center leaks value in the same place: the audio. Keyboard clatter, a colleague two desks away, traffic outside an agent's window at home — each one shaves accuracy off transcription, sentiment scoring, and every downstream model that reads the call.
CrystalSound is our answer to that. It sits in front of the rest of the stack and returns clean speech within a budget small enough that neither the agent nor the customer notices it is there.
CrystalSound runs as a streaming stage in front of transcription, so every downstream model sees enhanced audio.
Suppressing noise offline is a solved problem. Doing it in under 20 milliseconds, on a live call, is not.
The constraint that shapes every decision is latency. A model that looks ahead half a second produces beautiful output and an unusable conversation. Ours sees only a short window of future audio, which rules out most of the architectures that top the offline benchmarks.
The goal was never the highest score on a denoising benchmark. It was the highest word accuracy the transcription model could reach on real calls, at a latency nobody would complain about.
So we optimized against the metric we actually cared about. Instead of tuning for signal-to-noise ratio in isolation, we measured word error rate end to end, with the transcription model in the loop, on recordings from live queues rather than from a studio.
Conclusion
Enhancement is worth doing when it is judged by what it enables rather than by how it sounds. Every point of word error rate we remove flows straight into better summaries, better routing, and better scoring.
CrystalSound is deployed today across banking and insurance queues in Vietnam and Japan. The next step is pushing the same model onto the agent's device, so the noisy path disappears before it reaches the network at all.