
DNSMOS gets you most of the way there. It doesn’t tell you if the data is right for your model.
Written by DJ Jalali, Chief Technology Officer, with Geoff Bremner (Senior Audio Software Engineer, AI & Datasets) and Justin Hugh (Senior Product Manager, Voice Data)
Trusted by the world’s leading enterprises, AI technology, and software companies
Learn more about:
-
1
The limitations of DNSMOS
-
2
A four-step calibration framework you can apply to your own model
-
3
How Voices validates audio quality in producing custom datasets
Why ML teams read this
You’re building or fine-tuning a speech model and you’ve hit the same wall: DNSMOS tells you a score, not whether the data is right for your model. If you’re not calibrating DNSMOS and setting your own thresholds for acceptable data quality, you’re missing out on usable source data—increasing your training time and costs. This article will explain how to calibrate DNSMOS to your specific use case and maximize your source data.
What’s inside
- Where DNSMOS breaks down. We map what this means for your model.
- A calibration framework you can run this week. Four steps to set human-in-the-loop thresholds that match your specific use case to maximize your source data.
- The production side. How Voices captures and validates audio quality to produce custom speech datasets for ML engineers.