---

language: [zh, en] license: other license_name: fish-audio-community-license license_link: LICENSE base_model: fishaudio/drama-3 pipeline_tag: text-to-speech tags: [drama3, comfyui, audio, text-to-speech, voice-cloning, fp8, w4a8, t8star] library_name: comfyui

Drama-3 · T8 native models

English | 简体中文

Model files for Drama-3 T8 Studio & ComfyUI. Built with Fish Audio. Use the T8 native loader; these converted files are not standard Transformers checkpoints.

Start creating

The local white-and-pink Studio provides a versioned voice library, fixed casting, multiple takes, script editing, context-aware generation, recording segmentation and WAV/stem/subtitle delivery. ComfyUI uses the same core through native nodes, without a Studio HTTP bridge. See the code repository for installation, canvas workflows and version-bound validation.

The author supplies the complete Windows bundle. GitHub Release code ZIPs contain neither Python nor models. The bundle updater updates code while preserving runtime, models, user data and custom paths; it does not run pip.

Model update · 2026-10-29

FP8, W4A8 and both text heads now use the official 78e5ab7 weights. Download your selected main model and both matching heads together. Codec, tokenizers, ASR, aligner, VAD and speaker models are unchanged; existing encoded references remain compatible.

Use Studio/node 0.3.3 or later for the corresponding Windows loading and R2T2 compatibility fixes. Code auto-update preserves model files; update the model files separately. R2T2's multi-architecture CUDA runtime is distributed with the code release, not as a changed ASR model.

Exact source revisions, artifact hashes, conversion details and validation scope are in the new provenance record. FP8 expert weights/scales retain their source bytes; W4A8 is quantized from that matched FP8 conversion.

Files

Sizes are decimal GB/MB. Studio places these files directly under <bundle>/models/; the last column gives paths relative to ComfyUI.

Artifact	Size	ComfyUI destination / use
drama3-t8-fp8.safetensors	38.58 GB	models/diffusion_models/ — main model
drama3-t8-w4a8.safetensors	24.60 GB	Same folder — smaller main-model alternative
drama3-codec.safetensors	1.57 GB	models/vae/ — audio Codec
drama3-line-head.safetensors	8.46 KB	models/diffusion_models/ — compact Scene/control head
drama3-text-head.safetensors	1.03 GB	Same folder — full-head alternative/fallback
qwen3-asr-06b.safetensors / qwen3_asr_1.7b_bf16.safetensors	1.56 / 4.08 GB	models/asr/ — optional transcription
qwen3_forced_aligner_0.6b_bf16.safetensors	1.84 GB	models/asr/ — optional word timing
qwen-aligner/	1.84 GB model + tokenizer	models/asr/ — directory aligner alternative; download the complete folder
Confucius4-R2T2-GGUF/	2.18 GB pair	models/asr/ — Q8 model and matching projector
FireRedVAD-ONNX/	2.31 MB model	models/asr/ — R2T2 long-recording segmentation
campplus-zh-en.onnx	28.28 MB	Custom speaker-model path — optional CPU speaker grouping; Studio defaults to models/campplus-zh-en.onnx

Generation needs one main model, Codec and the selected/fallback head. ASR, alignment, VAD and speaker models are required only for their corresponding features. Reference-voice generation accepts clean audio with an accurate transcript; a supplied transcript needs no ASR. Keep the R2T2 model and projector together; its separately installed native runtime is also required.

All model paths are configurable, including absolute paths on other drives. Studio also supports custom voice-library and workspace locations. Code updates do not require downloading unchanged DRAMA weights. W6A8/INT8 conversion tools are available in the code repository; those checkpoints are not uploaded here.

The supported Windows stack is Torch 2.8 + CUDA 12.8, Transformers 4.57.1, tokenizers 0.22.2 and huggingface-hub 0.36.2. Memory and speed depend on the model, context and hardware; expert-cache size is not total memory use. Review generated speech before delivery. Existing release results do not certify every new feature or recording.

Sources and licenses

• DRAMA weights, heads and Codec retain the Fish Audio Community License v1.0, from fishaudio/drama-3. Redistribution and commercial use are conditional; the community commercial grant has a USD 500,000 annual-revenue threshold, and hosted services generally require a separate license. Review the full license before deployment.
• Qwen ASR/aligner, FireRedVAD and optional CAM++ use Apache-2.0. Preserve their licenses and notices: Qwen source/license, FireRedVAD, CAM++ source/license.
• Confucius weights retain the NetEase Youdao model license, including its downstream conditions and licensing thresholds. See the accompanying model notice. The Apache license of the inference code does not replace the weights' license.

Conversion sources and credits are recorded in current provenance, earlier provenance, upstream sources and notices. Existing artifact checksums are in model-manifest.json; the additional CAM++ checksum is in its notice. No upstream endorsement is implied.

T8star

Bilibili · YouTube · API · Gallery · Online apps · ComfyUI bundle · Hugging Face