Drama 3 focuses on one practical gap in speech generation: directing and iterating performances the way you direct actors. Rather than producing a single static TTS output, it treats speech generation as an interactive, editable production—supporting inline natural-language direction, multi-turn revisions, localized infilling edits, multi-character scenes, and chunked streaming for long recordings.
Key Capabilities
- Directed multi-character scenes: embed speaker tags and plain-language directions to produce scenes with up to 16 distinct voices in a single request, preserving timbre and roles across long dialogues.
- Editability and multi-turn revision: chain fixes and director notes so the model revises prior takes in context rather than re-generating from scratch, and perform localized infilling edits that keep surrounding acoustic transitions natural.
- Architecture and decoding: a 35B-parameter mixture-of-experts text backbone (≈3B active per token) paired with a codebook multi-token-prediction audio decoder; this design enables finer acoustic detail, multi-codebook decoding, and long-context causal decoding over frames.
- Production-ready streaming and throughput: engineered for low time-to-first-audio and high streaming throughput; supports deployment via SGLang-Omni and scales with vocoder placement to improve concurrent stream performance.
Who it's for and tradeoffs
Great fit if you produce scripted audio drama, podcasts, multi-character narration, or need iterative directed TTS workflows that require precise control over delivery and easy post-hoc edits. Look elsewhere if you need minimal-resource on-device inference (Drama 3 leverages a large MoE backbone and multi-GPU deployment patterns) or if you require an unrestricted commercial license—check the model license for usage constraints. The model improves expressiveness and editability at the cost of higher model size and deployment complexity compared with small single-GPU TTS systems.
