paksaz/VoxCPM2-Persian

VoxCPM2 Persian LoRA — Persian-only

This repository contains a Persian-only LoRA adapter for VoxCPM2. The base model is not included.

This is not a standalone TTS model. The adapter cannot be used on its own: download the Apache-2.0 VoxCPM2 base and load this adapter on top.

Checkpoint

Folder	Step	Text mode	Training data	val diff / total	Status
stageD_best_500	500	grapheme	clean Neyshekar + ParsVoice, 25,070 clips / 47.22 h	0.8346 / 0.9073	recommended

stageD_best_500 is the clean, audio-first checkpoint:

• full-sentence Persian data, 3–20 s
• trailing silence trimmed
• YODAS/YouTube ASR excluded
• no Pourmand
• 25,070 clips / 47.22 h (37,575 weighted rows)
• selected by audio validation, not loss alone

File checksum: SHA256SUMS.

Model details

• Base: VoxCPM2 (2B parameters), Apache-2.0.
• Adapter: LoRA r=256, alpha=256, LM + DiT + projection layers; ~1.69 GB.
• Training: 1× RTX 3090, bf16, batch 1 × gradient accumulation 16.
• Text side: Persian grapheme; no GE2PE and no phoneme input.

Quick start

pip install "voxcpm" huggingface_hub

from huggingface_hub import snapshot_download
from voxcpm import VoxCPM

repo = snapshot_download("paksaz/VoxCPM2-Persian")

model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM2",
    lora_weights_path=f"{repo}/stageD_best_500",
)

wav = model.generate(
    text="سلام، این یک آزمایش کوتاه است.",
    inference_timesteps=10,
    cfg_value=2.0,
    seed=42,
)

Evaluation

The base vs stageD_best_500 rows use Persian grapheme text only — no GE2PE.

Model	Persian s1 CER	Persian s2 CER	Persian s3 CER	Mean CER
base VoxCPM2 (no LoRA)	0.214	0.177	0.533	0.308
stageD_best_500	0.000	0.097	0.156	0.084

Persian sentences:

1. سلام، این یک آزمایش کوتاه است.
2. امروز هوا بسیار خوب است و می‌توانیم برای قدم زدن به پارک برویم.
3. کتاب من روی میز است و خانه ما نزدیک مدرسه است.

Samples: samples/base_0*_fa.wav and samples/stageD_500_0*_fa.wav. Raw rows: evaluation.json. All generations use cfg 2.0, 10 diffusion steps, seed 42.

This adapter is Persian-only. It is not intended for English, Arabic, or code-switching.

Training data and licenses

Dataset	License	Terms / attribution
Neyshekar (Sana Nazari)	CC0 1.0	No restrictions.
ParsVoice / Persian-Farsi-Speech (Mohammad J. Ranjbar)	CC BY 4.0	Attribution required; commercial use allowed.

Adapter license

The adapter weights in this repository are released under CC BY-NC 4.0: non-commercial use, with attribution. See LICENSE.

• The base model remains Apache-2.0; this repository does not redistribute it.
• The datasets are not redistributed here; their licenses apply to the data.
• If you need commercial use of the adapter, request permission from the author and check both dataset licenses above.
• You may not use the adapter without the VoxCPM2 base model, and you may not remove the attribution notices.

Intended use

• Persian text-to-speech research and personal projects.
• Auditioning/comparing Persian TTS adapters on top of VoxCPM2.
• Optional voice cloning with a short reference clip only with the speaker's consent.

Out of scope: English/Arabic/code-switching, commercial use (the adapter is CC BY-NC 4.0), impersonation, medical/legal/financial advice, or any use that violates the base model or dataset licenses. Do not present cloned speech as a real person's statement.

Limitations

• Persian only. No English, Arabic, or code-switch support.
• No guarantee of zero hallucination. stageD_best_500 was selected by audio validation: one short truncation in 10 held-out sentences, no long tails.
• VoxCPM2 is a 2B model. This is not a speed-optimized release.
• ASR metrics can hide pronunciation errors; listen to samples/ before deploying.

Citation / attribution

If you use this adapter, cite VoxCPM2 and the datasets above. At minimum:

VoxCPM2 Persian LoRA (2026). Persian-only LoRA adapter for openbmb/VoxCPM2.
Datasets: Neyshekar (CC0 1.0), ParsVoice (CC BY 4.0).