VoxCPM-Kernel/VoxCPM2-Turbo-distilled-DiT

VoxCPM2-Turbo: distilled LocDiT students

Faster replacements for the VoxCPM2 diffusion decoder (LocDiT). Each file holds only the LocDiT weights (feat_decoder.estimator.*). Load one on top of openbmb/VoxCPM2 or on top of VoxCPM2-Arabic-Dialects.

At a glance

Teacher	VoxCPM2's own sampler: 10 Euler steps, CFG 2.0
Changed weights	LocDiT only (feat_decoder.estimator.*)
Arabic dialects engine	student_k2_dmd.pt was its default until 2026-10-15. The default is now VoxCPM2-Arabic-Dialects-Distilled-DiT, distilled for its adapters
Trained with	VoxCPM2-Turbo

Files

File	Steps	What it is	Use it for
student_k2_dmd.pt	2	MSE + STFT, then distribution matching (DMD2-style)	Highest UTMOS of the few-step students
student_k2_stft2.pt	2	MSE + multi-resolution STFT	Comparison
student_k2.pt	2	MSE only	Comparison
student_k4.pt	4	MSE only	Comparison
student_gd9r2.pt	9	Guidance-distilled: one pass per step instead of two (CFG learned)	Recommended with fast mode
student_gd9pm.pt	9	Same, with a prefix-LM mask (cached conditioning K/V are exact)	Fast mode, alternative
lm_sparse24/sp24_ft.pt	–	2:4-pruned base LM after a recovery fine-tune (not a LocDiT)	Experimental
lm_sparse24/sp24_ft2.pt	–	Same, second recovery fine-tune with voice-cloning prompts	Experimental

Results

Few-step students (WER: Whisper large-v3-turbo; UTMOS over 3 seeds):

Decoder	Steps	WER	UTMOS
teacher (VoxCPM2 sampler)	10 + CFG	0.8–1.6%	3.83–3.89
student_k2_dmd.pt	2	1.19%	3.52
student_k2_stft2.pt	2	1.19%	3.36
student_k2.pt	2	0.27%	3.19
student_k4.pt	4	1.71%	3.33

Guidance-distilled students (100 held-out sentences × 2 seeds, fast mode, INT8 engine):

Decoder	WER	CER	Runaways	DNSMOS	UTMOS
teacher, fast mode	2.87%	1.52%	1	3.05	3.70
student_gd9r2.pt	2.38%	1.17%	0	3.14	3.94
student_gd9pm.pt	2.56%	1.28%	1	3.10	3.87

Median time to first audio with student_gd9r2.pt, 1× A10, fast mode, all requests sent at once:

Requests at once	1	8	32	128	256
teacher	10.3 ms	15.5 ms	38.4 ms	107.4 ms	175.7 ms
student_gd9r2.pt	10.1 ms	14.8 ms	33.0 ms	95.6 ms	156.4 ms

Usage

Arabic dialects engine (voxcpm2-arabic-dialects-turbo); its default is the Arabic student, so pass this file explicitly:

tts = await DialectTTS.load("VoxCPM-Kernel/VoxCPM2-Arabic-Dialects",
                            dit="VoxCPM-Kernel/VoxCPM2-Turbo-distilled-DiT:student_k2_dmd.pt")

VoxCPM2-Turbo:

eng = Engine("openbmb/VoxCPM2", precision="int8",
             smooth_alpha={"qkv": 0.7, "o": 0.7, "gate_up": 0.7, "down": 0.7},
             student="hf://VoxCPM-Kernel/VoxCPM2-Turbo-distilled-DiT/student_k2_dmd.pt")

For student_gd9*.pt, call VoxCPM2.enable_fast_mode() first (server: --fast).

License

Apache-2.0, as VoxCPM2.