Audio Training Data: DNSMOS, Reverb, and the Grey Zone
DNSMOS gets you most of the way to audio quality. It doesn’t explicitly measure reverb, even when it’s baked into your data. You're sourcing or reviewing speech data for a TTS, voice cloning, or ASR model, and it clears your DNSMOS threshold. But a DNSMOS score in the 3–4 range, the grey zone where most real-world data lives, can still hide reverb: the reflected sound energy that lingers after a speaker stops talking. For voice cloning, that reverb gets baked into the voice characteristics as part of the voice and reproduced in every sentence the model generates.
We’ll be sitting down with Geoff Bremner, Senior Audio Software Engineer at Voices, to discuss:
Where DNSMOS breaks down, and how to evaluate whether the 3–4 grey zone contains reclaimable data for your use case
What reverb actually is: RT60, reflection tails, and why they quietly erode perceived audio quality
A live audio example, so you can compare and hear the difference
Who should attend: ML and data engineering teams sourcing or evaluating audio quality for TTS, voice cloning, or ASR models.
About our speaker
Geoff Bremner works at the intersection of audio and software engineering—building the voice data pipelines that power AI systems for leading technology companies.
He holds a computer science degree from the University of Waterloo and trained in audio production at OIART, bringing a rare dual background: he's produced 500+ podcasts and 5+ studio albums, with thousands of hours spent recording, editing, and evaluating spoken-word content. Today, he applies that expertise to designing scalable workflows for recording, quality control, post-production, and data management—turning voice talent's contributions into high-quality training data for marketing-leading AI technology.
About AlphaSignal
AlphaSignal is a technical AI news company trusted by 300,000+ engineers, researchers, and technical founders. AlphaSignal cuts through the noise to bring you the tools, papers, and breakthroughs that actually matter - no fluff, just what you need to stay ahead in AI.
Join AlphaSignal for free