VoxCPM-Kernel/VoxCPM2-Arabic-Dialects

VoxCPM2 Arabic Dialects

VoxCPM2 with 11 LoRA adapters, one per dialect. A server keeps every adapter loaded and picks one per request, so one batch can mix dialects. Use it for fast voice cloning in these dialects; for one model that covers 17 dialects, see Teacher-v3.

At a glance

Base model	openbmb/VoxCPM2 (2B parameters, 48 kHz), weights unchanged
Trained	11 LoRA adapters: rank 64, alpha 128, on the attention (q/k/v/o) of the language models and the LocDiT
Dialects	10 Arabic adapters + Hindi-English (table below)
Size	6.55 GB: base 4.96 GB + 145 MB per adapter
License	Apache-2.0
Best for	Serving many dialects at once with the Turbo engine
Fast-mode LocDiT	VoxCPM2-Arabic-Dialects-Distilled-DiT: 2 steps, distilled from these adapters (the engine's default)

Adapters

adapter	dialect	training step
ab	Jordanian (a second Jordanian adapter; its weights differ from jordanian)	67,500
egyptian	Egyptian	49,000
emirati	Emirati	96,000
hejazi	Hejazi	72,000
hindi_english	Hindi-English code-switched	40,500
iraqi	Iraqi	28,000
jordanian	Jordanian	149,000
najdi	Najdi	77,500
palestinian	Palestinian	76,500
qatari	Qatari	53,000
saudi	Saudi	78,000

Files

path	what it is
model.safetensors, audiovae.pth, config.json, tokenizer files	the base VoxCPM2 checkpoint, same weights as openbmb/VoxCPM2
adapters/<name>/	one dialect's LoRA: lora_weights.safetensors (fp32) and lora_config.json
adapters.json	index of the adapters: path, description, training step

Results

One NVIDIA A10 with the Turbo engine, cloning the 32 voices of VoxCPM-Kernel/Arabic-Hindi-Reference-Voices, each with its own dialect's adapter.

requests at once	first audio, avg	first audio, p95	whole sentence, avg
1	23 ms	24 ms	0.35 s
16	26 ms	27 ms	0.49 s
64	40 ms	48 ms	0.88 s
128	60 ms	60 ms	1.37 s
200	92 ms	92 ms	1.94 s
256	115 ms	117 ms	2.39 s

mode	LocDiT steps	first audio p95, 200 requests	CER	WER	UTMOS
fast (default)	2, distilled for these adapters	92 ms	5.3 %	16.1 %	2.14
quality (dit=None)	10, with CFG	270 ms	5.3 %	15.5 %	2.40

• First audio: time until the first audio chunk is ready. CER / WER: character / word error rate of a speech recognizer's transcript (better of Whisper large-v3 and Cohere Transcribe Arabic); lower is better.
• UTMOS: predicted naturalness, 1-5; higher is better.
• Quality columns: a held-out evaluation set, 17 dialects × 40 sentences × 2 seeds (1,320 clips), each voice cloned from another clip of its speaker, int8. Per dialect, and Hindi-English: the distilled LocDiT's card.

Usage

Install the Turbo engine (./setup.sh), then:

import asyncio
from nanovllm_voxcpm import DialectTTS

async def main():
    tts = await DialectTTS.load("VoxCPM-Kernel/VoxCPM2-Arabic-Dialects")   # fast mode: the 2-step Arabic LocDiT
    await tts.add_voice("ahmed", "ahmed.wav")
    audio = await tts.synthesize("هلا والله، كيف حالك اليوم؟", dialect="najdi", voice="ahmed")  # 48 kHz
    await tts.close()

if __name__ == "__main__":
    asyncio.run(main())

HTTP server (every adapter is registered; a request picks one with "dialect"):

python -m nanovllm_voxcpm.entrypoints.openai_server --model VoxCPM-Kernel/VoxCPM2-Arabic-Dialects \
    --dit VoxCPM-Kernel/VoxCPM2-Arabic-Dialects-Distilled-DiT:student_k2_dmd.pt --port 8092
curl localhost:8092/v1/audio/speech -H 'Content-Type: application/json' \
    -d '{"input": "هلا والله، كيف حالك اليوم؟", "dialect": "saudi", "response_format": "wav"}' > out.wav

Training data and license

The adapters are reorganised from an earlier upload; their training data is not documented here. License: Apache-2.0, like the base model VoxCPM2.