VoxCPM-Kernel/VoxCPM2-Arabic-Dialects-Samples

VoxCPM2 Arabic Dialects: voice-cloned samples, optimized vs original

50 requests to VoxCPM-Kernel/VoxCPM2-Arabic-Dialects (VoxCPM2 + 11 dialect LoRA adapters), each cloning one of 5 Arabic reference voices. Every row holds the same request generated twice, to listen to side by side:

• optimized_audio: the optimized engine (Speech-Next/voxcpm2-arabic-dialects-turbo), as one of 200 requests sent at the same time (p95 time to first audio ~96 ms on one NVIDIA A10);
• original_audio: the original VoxCPM2 implementation with no optimizations, one request at a time.

Each with its Whisper large-v3 transcript, WER, CER, UTMOS, DNSMOS and speaker similarity to the reference voice.

Optimized vs original (these 50 requests)

	optimized	original
speed	first audio in ~95 ms, 200 requests at once	3.6 s per utterance, one at a time
WER	14.3 %	20.0 %
CER (mean / median)	7.1 % / 2.7 %	9.8 % / 2.6 %
UTMOS	2.59	2.79
DNSMOS OVRL	2.88	2.91
speaker similarity to the reference	0.923	0.927

The original's mean CER is raised by one utterance (row 4, hindi_english) that drifted into other words; without it the original is the more accurate one: CER 5.8 % against 7.1 %. Per row, the optimized version has the lower CER in 15 rows, the original in 14, the rest are equal. The original also sounds slightly cleaner (UTMOS +0.2); voice similarity is the same.

Per dialect:

dialect	n	WER optimized	WER original	CER optimized	CER original
ab	5	7.9 %	10.4 %	2.1 %	1.9 %
egyptian	5	2.9 %	5.4 %	1.0 %	1.5 %
emirati	5	11.1 %	22.9 %	5.1 %	10.3 %
hejazi	5	6.2 %	9.0 %	1.2 %	1.9 %
hindi_english	5	39.4 %	70.9 %	17.5 %	52.2 %
iraqi	5	9.5 %	28.2 %	5.5 %	9.5 %
jordanian	4	21.3 %	20.6 %	22.0 %	4.4 %
najdi	4	11.8 %	9.4 %	1.7 %	1.9 %
palestinian	4	12.1 %	6.3 %	3.4 %	1.2 %
qatari	4	10.3 %	7.1 %	7.8 %	7.1 %
saudi	4	27.6 %	23.1 %	13.6 %	11.6 %

Time to first audio, 200 simultaneous requests (NVIDIA A10)

	p95
base inference (VoxCPM-Kernel/voxcpm-inference), no cloning	2174 ms
base inference, voice cloning	4464 ms
optimized engine, no cloning	96 ms
optimized engine, voice cloning	96 ms

With voice cloning, by the number of requests sent at the same time (p95):

simultaneous requests	1	8	16	64	128	256
base inference	93 ms	188 ms	364 ms	1455 ms	2838 ms	5708 ms
optimized engine	21 ms	24 ms	27 ms	43 ms	67 ms	123 ms

How they were made

• Request i: sentence i of ar_dialects.txt, adapter i % 11, voice i % 5 (reference cloning, the voice's timbre from its clip), seed 2000 + i, cfg 2.0.
• Optimized: distilled 2-step LocDiT, int8 layers, the 11 adapters merged into int8 weights in one engine, head start (the first 3 patches from the first <= 8 tokens of the text); the first 50 of a burst of 200.
• Original: voxcpm 2.0.3 (openbmb, PyTorch) with the request's adapter loaded, optimize=False, bf16, 10 diffusion steps, its defaults otherwise (including its retry of bad outputs).
• Scoring: Whisper large-v3 (openai/whisper-large-v3 via faster-whisper, language ar, greedy), WER and CER after Arabic normalization; UTMOS22; DNSMOS P.835; WavLM-SV cosine to the reference clip.

Files

path	content
data/train-00000-of-00001.parquet	the rows: both audios (48 kHz) and every score
audio/, audio_original/	the same audio as WAV files
references/<voice>.wav, .json	the 5 reference clips (16 kHz), transcripts and FLEURS ids
metadata.csv, summary.json	the rows without audio; the aggregates

License and attribution

The reference clips are from FLEURS (Arabic, ar_eg train split), CC BY 4.0, Conneau et al., "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech", 2022. The generated audio is released under CC BY 4.0 as well. The model is VoxCPM2 (Apache-2.0) with the dialect adapters of VoxCPM-Kernel/VoxCPM2-Arabic-Dialects.