OGG TTS: open-source 0.75B expressive multilingual TTS with block diffusion and emotion tags
solyarisoftware · x · 2026-11-07
AlphaCEP released a preview of OGG TTS on Hugging Face, a 0.75B-parameter zero-shot multilingual TTS model.
- Trained on 8,200 hours of studio-quality speech across 8 languages (English, Polish, Spanish, German, Russian, French, Portuguese, Italian)
- Backbone initialized from OmniVoice (Qwen3-0.6B, masked discrete diffusion); generates Fish Audio 44.1kHz codec tokens via block diffusion — each 0.74s block is produced by parallel iterative unmasking, and audio streams as soon as the first block is ready; an axial multi-token prediction decoder fills in acoustic detail
- Naturally expressive by default, 24 inline emotion/delivery tags, zero-shot voice cloning from seconds of reference audio
- Currently a 120k-step checkpoint of a 400k-step pre-training run; OGG is the first of three architectures (OGG/ROGG/POGG) compared in an upcoming technical report
More from Multimodal
- Midjourney 8.2 test: stunning Munch-inspired expressionist artwork — LudovicCreator · 2026-11-07
- Midjourney style code --sref 4173596805: cinematic grainy night scenes of the Colosseum — sergeantsref · 2026-11-07
- How I built a rhythm racing game with Claude Opus 5.5, six skills, and a voice feedback loop — Groowl · 2026-11-07
- Stylized 2D-to-3D character conversion preserves every design detail — azed_ai · 2026-11-07
- Doki lets you make videos with no timeline — just write the whole thing as one story — thesephist · 2026-11-07
- hyperflow-h3-flash video model updated: sampling in just 4, 6, or 8 steps — Apprehensive_Sky892 · 2026-11-07