VoxCPM-Kernel/VoxCPM2-Arabic-Dialects-Latency-Samples

VoxCPM2 Arabic Dialects: time-to-first-audio samples

Five clips from a latency test: 200 voice-cloning requests sent at the same instant to one NVIDIA A10 GPU. Listen to hear what the engine produces under that load; data/metadata.csv has every number per clip.

At a glance

What	5 generated clips and the reference voice they clone
Load	200 requests at the same instant, one NVIDIA A10 (24 GB)
Time to first audio	95.5-96.8 ms for these 5 clips; 96.8 ms p95 over all 200 requests
Dialect adapters	egyptian, iraqi, najdi, palestinian, saudi
Model	VoxCPM-Kernel/VoxCPM2-Arabic-Dialects (VoxCPM2 + 11 dialect LoRA adapters)
Engine	Speech-Next/voxcpm2-arabic-dialects-turbo
Measured	2026-10-11
License	CC BY 4.0

Time to first audio (TTFA) is the time from sending a request to receiving its first audio chunk, which is 40 ms of audio. The rest of the sentence streams in faster than it plays: a 3-4.5 s sentence is fully generated about 2 s after the request.

Samples

file	sentence	words	TTFA	audio	whole sentence ready	CER	WER
001_egyptian.wav	المكيف في الغرفة ما يبرد زين.	6	95.5 ms	3.04 s	1.94 s	0.0 %	0.0 %
005_iraqi.wav	كيفك؟ شو عم تعمل هالأيام؟ صرلنا زمان ما حكينا.	9	95.9 ms	4.48 s	2.39 s	0.0 %	0.0 %
007_najdi.wav	السيارة تبي لها بنزين، خلنا نوقف عند أقرب محطة.	9	96.3 ms	4.00 s	2.28 s	0.0 %	0.0 %
008_palestinian.wav	ولدي بدأ يتعلم السباحة هالصيف.	5	96.4 ms	3.52 s	2.13 s	3.4 %	20.0 %
010_saudi.wav	أبي أحجز طاولة لأربعة أشخاص الساعة تسع.	7	96.8 ms	3.52 s	2.13 s	23.7 %	42.9 %

CER / WER: Whisper large-v3 transcript against the sentence, after Arabic normalization. All 200 outputs of the burst: CER 6.0 %, WER 14.7 %, UTMOS 2.95, speaker similarity 0.92.

Files

path	content
data/*.wav	the 5 generated clips
data/metadata.csv	one row per clip: file_name, request, text, words, adapter, concurrent_requests, ttfa_ms, burst_ttfa_p95_ms, audio_s, all_audio_ready_s (request to last chunk), chunks, stall_s (playback wait after start), whisper_large_v3, cer_pct, wer_pct, speaker_similarity (WavLM-SV), utmos (1-5)
reference/ar00.wav	the voice every request clones

How it was measured

Engine settings	int8 W8A8 (SmoothQuant), 2-step distilled LocDiT, CUDA graphs; the first 40 ms frame is streamed ahead of the rest
Voice cloning	from the reference audio only, without its transcript
Requests	request i uses adapter i mod 11 and one of 192 short dialect sentences, so the adapter does not always match the sentence (the iraqi request reads a Levantine sentence)
Bursts	1 warm-up, then 3 measured: p95 TTFA 96.8, 95.4 and 94.6 ms; the clips are from the first
Selection	the first request of each of the 5 adapters, not chosen by quality

Load

from datasets import load_dataset

ds = load_dataset("VoxCPM-Kernel/VoxCPM2-Arabic-Dialects-Latency-Samples", split="train")
print(ds[0]["text"], ds[0]["ttfa_ms"])

License

CC BY 4.0. The reference voice reference/ar00.wav is a google/fleurs clip (ar_eg train, id 1480, CC BY 4.0). The audio was generated with VoxCPM2 (Apache-2.0).