End-to-end Japanese multi-speaker conversation model trained on 130k+ hours runs on consumer GPUs

kastnerkyle · x · 2026-11-14

Developer MystiqCaleid released an end-to-end conversation generation model for Japanese, tackling two underexplored problems: generating multi-speaker conversations directly, and scaling on a low budget. The architecture is efficient enough to train on consumer GPUs, using over 130k hours of Japanese speech data.

Original post →

End-to-end Japanese multi-speaker conversation model trained on 130k+ hours runs on consumer GPUs

More from Multimodal