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

VoxCPM2 Arabic Dialects: TTFA samples

Five outputs from one latency test: 200 voice-cloning requests sent at the same instant to one NVIDIA A10, each asking for a different Arabic dialect sentence. In this burst the time to first audio (TTFA) was 95.5-96.8 ms for the five samples and 96.8 ms at p95 over all 200 requests.

What TTFA measures

TTFA is the time from sending a request to receiving its first audio chunk. That chunk is only 40 ms of audio, the start of the first syllable. The rest of the sentence keeps streaming in after it, 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. Over all 200 outputs of the burst: CER 6.0 %, WER 14.7 %, UTMOS 2.95, WavLM-SV speaker similarity to the reference 0.92.

data/metadata.csv has every number per sample:

column	meaning
request	index of the request in the burst
adapter	dialect adapter (LoRA) used
concurrent_requests	requests sent at the same instant
ttfa_ms	time to first audio of this request
burst_ttfa_p95_ms	p95 TTFA over the burst
audio_s	length of the generated audio
all_audio_ready_s	time from the request to its last audio chunk
chunks	audio chunks streamed
stall_s	time playback would have waited for audio after it started
whisper_large_v3	Whisper large-v3 transcript
cer_pct, wer_pct	character / word error rate
speaker_similarity	WavLM-SV similarity to the reference voice
utmos	UTMOS predicted naturalness (1-5)

Setup

• Model: VoxCPM2 with 11 Arabic dialect LoRA adapters.
• Engine: a fork of nano-vllm-voxcpm on one NVIDIA A10 (24 GB). It runs with int8 W8A8 (SmoothQuant), a 2-step distilled LocDiT and CUDA graphs, and streams the first 40 ms VAE frame ahead of the rest of the patch.
• Voice cloning: every request clones the voice in reference/ar00.wav from reference audio only, without its transcript.
• Requests: request i uses adapter i mod 11. The sentences come from a set of 192 short sentences in several dialects (7 words on average), so a request's adapter does not always match its sentence's dialect. For example, the iraqi request reads a Levantine sentence.
• Burst: one unmeasured warm-up burst, then 3 measured bursts with the same 200 requests. The p95 TTFA of the three bursts was 96.8, 95.4 and 94.6 ms. The audio and numbers here are from the first measured burst.
• Selection: the five samples are the first requests of the egyptian, iraqi, najdi, palestinian and saudi adapters. They were not chosen by quality: the saudi sample has a CER of 23.7 %.
• Date: measured on 2026-10-11.

License

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