VoxCPM-Kernel/VoxCPM2-Arabic-Train-v2

VoxCPM2 Arabic Train-v2: 17-dialect TTS training set

The training set of the Arabic VoxCPM2 teachers: 135 h of speech in 17 Arabic dialects, stored as AudioVAE latents with a manifest. Download it and train with Speech-Next/voxcpm2-arabic-training.

At a glance

Size	47,574 segments, 135.3 h, 17 dialects
Real speech	102.9 h: lahgtna 97.3 h, SADA 5.6 h
Synthetic speech	32.4 h (VoxCPM2-Arabic-Dialects adapters and Habibi-TTS)
Splits	train 47,326 clips / 134.67 h; validation 248 clips / 0.67 h
Voices	2,074 real speakers (at most 2.45 h each), 451 synthetic voices
Format	AudioVAE latents (fp16, 25 Hz × 64) and a JSONL manifest; no waveforms
Trained on it	Teacher-v2; Teacher-v3 (same clips, reweighted)
License	CC BY-NC-SA 4.0, non-commercial

Hours per dialect (train split)

dialect	lahgtna	SADA	synthetic	total
bahraini	23.03			23.03
sudanese	11.89			11.89
yemeni	11.72	0.01		11.73
egyptian	10.65	0.04		10.69
palestinian	9.49			9.49
najdi		3.16	6.32	9.48
syrian	9.46			9.46
hejazi		0.81	7.28	8.09
saudi	1.65	0.35	4.49	6.50
lebanese	4.89		1.25	6.14
algerian	5.18		0.79	5.97
moroccan	3.17		1.34	4.51
tunisian	2.37		2.01	4.38
iraqi	1.62		2.49	4.10
gulf		0.55	2.98	3.54
libyan	1.85		1.49	3.34
msa		0.45	1.89	2.34
all	96.97	5.38	32.33	134.67

Contents

path	content
manifest.jsonl	this set, one row per segment
stats.json	clips and hours per split, source and dialect; the filter funnel; quality pass rates
{lahgtna,sada,synth}/latents/	AudioVAE latent shards of this set's clips only
{lahgtna,sada,synth}/manifest/	the corpora's full clip manifests
quality/scores.jsonl	Cohere CER and DNSMOS of 83k real candidate clips
eval/v2/heldout.json	evaluation speakers, clips and texts kept out of training

Main manifest fields: uid, split (train / val), source (lahgtna / sada / synth), speaker, dialect, duration_s, frames, text (the cleaned transcript), scores (cohere_cer, dnsmos_ovrl), latents (the shard), license; synthetic rows also generator, voice.

How clips were selected

filter	real (lahgtna, SADA)	synthetic
length	1-20 s	1-20 s
transcript check (Cohere Transcribe CER)	≤ 10 %	≤ 8 %
audio quality (DNSMOS OVRL)	≥ 3.0	≥ 3.0
speaking rate	2-35 characters/s	5-25 characters/s
per speaker	≤ 3 h	–
left out	evaluation speakers, clips and texts; SADA's test split; texts with digits	evaluation texts; texts with digits

The validation split is SADA's validation split plus a 0.4 % random sample; it is never trained on. Of 592 h of lahgtna and 53 h of SADA, 97.3 h and 5.6 h pass: DNSMOS removes most of SADA (TV drama with music and reverberation). stats.json has the full funnel.

Download and train

hf download VoxCPM-Kernel/VoxCPM2-Arabic-Train-v2 --repo-type dataset --local-dir /workspace/data/arabic
python -m arabic_voxcpm.train.train --out runs/my-teacher --steps 3000 \
    --corpora lahgtna,sada,synth --quality /workspace/data/arabic/quality/scores.jsonl \
    --heldout /workspace/data/arabic/eval/v2/heldout.json \
    --lora '{"train_modules": ["base_lm"], "enable_dit": false, "enable_proj": false}'

The latents were re-encoded from the pinned source revisions on 2026-10-14 with TF32 off. They are not bit-identical to the ones Teacher-v2 was trained on, but differ only at the encoder's own noise level (about 6e-4 relative).

Sources and licenses

source	license
Rabe3/lahgtna-arabic-tts-24khz @ f3582bec	none declared
MohamedRashad/SADA22 @ 094fe2c0	CC BY-NC-SA 4.0 (SDAIA, Saudi Broadcasting Authority)
synthetic: VoxCPM2-Arabic-Dialects adapters or Habibi-TTS, corpus texts, VoxTube or corpus voices	VoxTube: CC BY-NC-SA 4.0

Treat the whole set as non-commercial (CC BY-NC-SA 4.0 or stricter).