skysay/drama3-emo-w4

Skysay Drama 3 Emo-W4: an emotion-calibrated 4-bit Drama 3 for ComfyUI

Built with Fish Audio. This is a quantized derivative of fishaudio/drama-3 and is covered by the Fish Audio Community License (see LICENSE and NOTICE).

What it is. The routed experts are mixed-precision: W4A8 by default, with W6A8 on layer 0, the one layer whose MoE output breaks under 4-bit on the emotional calibration speech (relative output error 1.72 at 4-bit vs 0.0015 at 6-bit; every other layer is 0.003–0.013). Dense layers are bf16. It loads in ComfyUI-Drama3 the same way as its other files.

file	size	peak VRAM	experts in RAM (8 GB cache)
drama3_emo_w4.safetensors	24.8	27.7	not supported yet (mixed banks; use experts in VRAM)

Why

Plain 4-bit quantization keeps WER but can flatten crying and laughter. This file is calibrated on emotional speech, so the expressive categories survive.

Results: 30 lines × 3 voices × 7 seeds = 630 takes per variant, seed-matched, 95% bootstrap CI

variant	size GB	WER	fail	sob fail	spk-sim	emotion score
FP8 (pack)	38.6	0.043	0.049	0.190	0.519	0.124
W4A8 (pack)	24.6	0.038	0.049	0.190	0.523	0.119
Emo-W4 (this)	24.8	0.031	0.039	0.143	0.533	0.118

"fail" means silence, a stall cut, truncation or babble. WER is Whisper large-v3 with normalization. The emotion score is the AST AudioSet probability for the target event. Full method: github.com/skyzer/drama3-quant-bench.

Sob, 14 seeds (168 takes per variant): Emo-W4 0.149, FP8 0.202, W4A8 0.226. Fail Δ vs W4A8 is −0.077 [−0.137, −0.018] (significant); vs FP8 it is −0.054 [−0.149, +0.024] (CI includes 0). Overall (630): fail Δ vs FP8 is −0.010 [−0.033, +0.012] and WER Δ is −0.013 [−0.032, +0.005]. Both are non-inferior; not claimed better. Trade-offs: the cry-event score on sob (0.048) is between W4A8 (0.034) and FP8 (0.063). Soblaugh fail is 0.127 (W4A8 0.119, FP8 0.143).

Listen (LJ Speech public-domain reference voice; same seed for every file)

line	FP8	W4A8	Emo-W4
[sobbing] I'm sorry, I just, I really thought he would come home this time.
[sobbing and laughing at the same time] I'm so happy…
[laughing] Oh no, you did not just wear that to the wedding!
[screaming] Get out of the way…
[whispers] Don't move…
The quarterly report is ready…
Every file uses seed 11 and the same seed per row, no cherry-picking. Plain links: samples/

Talking-video demo (Drama 3 Emo-W4 → MiniMax H3 Ref2VA; synthetic face generated with Z-Image Turbo, LJ voice)

The workflow is at github.com/skysay-ai/drama3-h3-talking. MiniMax H3 has its own license, including a territory clause.

Calibration recipe

• 26 new lines (not in the test set): 6 sob, 5 laugh, 4 scream, 4 whisper, 3 sob-laugh and 4 neutral. Each line was generated with a non-test reference voice and with no reference at all. The LJ test voice was never used for calibration.
• MoE inputs were captured on the FP8 model, with neutral tokens down-weighted to 0.3. For each layer we computed the relative MoE-output error at 4 and 6 bits, then promoted the outlier layer (L0) to W6A8. A 12-layer W6 variant with int8 dense layers (24.66 GB) did worse on sob (0.179) and was dropped.
• Tools: github.com/skyzer/drama3-quant-bench/b2 (GPLv3, built on the ComfyUI-Drama3 converter).

Limits

• There are only 4 test lines per emotion category, so the confidence intervals are wide. The experts-in-RAM offload in ComfyUI-Drama3 assumes uniform banks and does not load this file yet.
• The automatic emotion metrics are a proxy for human listening.
• MLX/Mac users: see github.com/skyzer/ComfyUI-Drama3-MLX.

License

Fish Audio Community License. Commercial use is free only below USD 500k annual revenue, and offering this model as a hosted TTS or voice-cloning service needs a separate license from Fish Audio. Don't clone voices without consent.