Dessa's dialog-only voice training isn't new: dGSLM did it in 2022 with just 2k hours

rdesh26 · x · 2026-10-24

Responding to the buzz around Dessa being trained only on human conversations (no text, no pretrained models), the author pushes back on two fronts:
Not a first: dGSLM already did dialog-only training in 2022 with just 2k hours of Fisher two-channel data; vendors now sell thousands of hours of high-quality 2-channel dialog off the shelf.
Different != smart: Writing took humans hundreds of thousands of years to invent and 5k years to optimize for efficient knowledge storage. Letting a model learn from text first, then spoken dialog (like Moshi and basically every modern S2S model), cuts training cost 10x. Dialog-only training was a limitation of the era, not a deliberate advantage.
The author concedes the natural chit-chat quality is impressive but questions why anyone would skip text pretraining today.

More from Models