VoxCPM2 Arabic Dialects: optimized vs original engine, voice-cloned samples
50 requests to VoxCPM-Kernel/VoxCPM2-Arabic-Dialects,
each generated twice: by the optimized engine under heavy load and by the original VoxCPM2 code, one at a time. Every
row holds both audios and their scores, to compare side by side.
At a glance
|
|
| What |
50 rows: same text, voice and seed, generated by two engines |
optimized_audio |
Speech-Next/voxcpm2-arabic-dialects-turbo, one of 200 requests sent at once, one NVIDIA A10 |
original_audio |
the original voxcpm 2.0.3 code (bf16, 10 diffusion steps), one request at a time |
| Dialect adapters |
11: ab, egyptian, emirati, hejazi, hindi_english, iraqi, jordanian, najdi, palestinian, qatari, saudi |
| Voices |
5 Arabic reference clips from FLEURS |
| License |
CC BY 4.0 |
Results (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 |
One original utterance (row 4, hindi_english) drifted into other words and raises its mean CER; without it the
original is slightly more accurate (CER 5.8 % vs 7.1 %) and slightly cleaner (UTMOS +0.2). Voice similarity is the
same.
| 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 with voice cloning, p95 (NVIDIA A10)
Columns
| column |
content |
id, dialect, voice, text, seed, clone_mode |
the request |
optimized_audio, original_audio |
the two generations (48 kHz) |
optimized_whisper, original_whisper |
Whisper large-v3 transcripts |
*_wer, *_cer |
error rates after Arabic normalization |
*_utmos, *_dnsmos, *_voice_sim |
UTMOS22, DNSMOS P.835 OVRL, WavLM-SV similarity to the reference |
optimized_ttfa_ms, original_generation_s |
time to first audio; time to generate the whole utterance |
Other files: audio/ and audio_original/ (the same audio as WAV), references/ (the 5 reference clips with their
transcripts and FLEURS ids), metadata.csv (rows without audio), summary.json (aggregates).
How they were made
- Request i: sentence i, adapter i mod 11, voice i mod 5, seed 2000 + i, CFG 2.0.
- Optimized: 2-step distilled LocDiT, int8 layers with the 11 adapters merged into one engine; the first 50 of a burst of 200.
- Original:
voxcpm 2.0.3 with the request's adapter, optimize=False, its defaults otherwise (including its retry of bad outputs).
Load
from datasets import load_dataset
ds = load_dataset("VoxCPM-Kernel/VoxCPM2-Arabic-Dialects-Samples", split="train")
print(ds[0]["text"], ds[0]["optimized_cer"], ds[0]["original_cer"])
License
CC BY 4.0. The reference clips are from FLEURS (ar_eg train,
CC BY 4.0; Conneau et al., 2022). The model is VoxCPM2 (Apache-2.0) with the adapters of VoxCPM2-Arabic-Dialects.
