MOSS-TTS voice-acting demo grids (DramaBox reinterpretations, LoRA merges, VoiceNet & character voices)

MOSS-VA-v2 conditioning manual: 40 emotions (best prompt + LoRA merge %, reward, correlations) — MD skill + HTML

Evolutionary edge-case search: extreme vocalizations per emotion LoRA

Evolutionary maximization of all 40 emotions (10 generations, mean-of-8)

VoiceNet-dimension LoRA evolution: 50% vs 100% dose + optimal prompting (10 generations)

Emotion prompt x LoRA-strength study: 40 emotions x 6 neutral sentences x {neutral, prompt, LoRA 50/100/150%}

User interaction data is mostly opaque and saved in big tech companies (Claude, OpenAI, Google). This project tries to engage the public to use v1-compatible endpoints and frontend chats to donate its user interaction data for the development of future open frontier models.

LAION MOSS Local-Transformer 1.5 voice-acting model (4.55B, 48 kHz) — inference, fast-inference guide, best-of-64 audio demos

Self-contained DramaBox voice acting pipeline: VoiceNet taxonomy, multi-GPU prompt generation, TTS synthesis, and audio refinement