Duplex Model Training Whitepaper
Open-source and synthetic data are the default route to training conversational AI. Yet, neither teaches a model to tell a back channel from an interruption.
Most conversational datasets are half-duplex, with overlap that's either missing or unannotated. Even Fisher, the most widely used full-duplex corpus, only captures overlapped speech for about 14% of speech duration. You can stitch an interruption into a transcript after the fact, but that won't reproduce the timing of real conversation.
This whitepaper covers the three requirements for duplex-ready training data, where open-source and synthetic datasets fall short, and how we produce custom, role-differentiated datasets built specifically for duplex training.
Written by DJ Jalali (Chief Technology Officer), with Sulayman Goolamnobee (Product Manager, AI Voice).
Request access above to download the full whitepaper.
Once granted access, download it from the Files and versions tab above, or go directly to the PDF file.
License
Use of this whitepaper is governed by the Voices.com Whitepaper License.
- Downloads last month
- 31