Fish Audio open-sources Drama 3: 35B MoE TTS with per-word tone control
ying11231 · x · 2026-10-14
Fish Audio released Drama 3, an open-weight TTS model billed as its most emotive and controllable yet, with SGLang offering day-0 support.
- Architecture: 35B-param MoE text backbone (3B active/token) with hybrid full+linear attention; Codebook-MTP audio decoder decoding 10 codebooks causally; FP8 experts with bf16 elsewhere
- Capabilities: up to 16 speaker tags per text, voice cloning from reference audio, natural-language control down to single-word tone and mid-sentence voice switching
- Serving: SGLang-Omni serves multi-speaker TTS and streaming audio via /v1/audio/speech, with 89ms median time-to-first-audio on one H200
- The company recently raised a $52M seed round and claims 15M+ users
Related event: Fish Audio Open-Sources Drama 3 TTS Model, Discloses $52M Seed Round(4 posts)→
More from Infra
- AMAT investor QA: inference-specific accelerators may not need 12-hi HBM stacks — BenBajarin · 2026-10-14
- Liquid AI hits 21k+ output tokens/s on NVIDIA B300 at high concurrency — JosephJacks_ · 2026-10-14
- Disaggregated Quantization lifts 1-bit Qwen3.8-27B from 29% to 61.5% on MMLU-Pro — udmrzn · 2026-10-14
- HINT4 targets a hardware-invariant NVFP4: 4-bit inference via int8 tensor cores on Ampere — secemp9 · 2026-10-14
- VigenFlow runs FLUX and Z-Image locally on AMD Ryzen AI NPU, no GPU needed — Kelsey_test · 2026-10-14
- RTX 5090 owner seeks free ComfyUI motion transfer workflows to avoid paid Higgsfield — Quirky_Fig3927 · 2026-10-14