Megjegyzés a magyar változathoz:
VoxCPM2_hun_alpha_2026.05.04
A VoxCPM2_hun_alpha_2026.05.04 egy korai, kísérleti magyar LoRA-alapú finomhangolás.
A teljes finomhangoláshoz legalább 48 GB VRAM szükséges, ami jelenleg saját hardveren nem áll rendelkezésemre. Emiatt ez a verzió egyelőre nem teljes értékű tréning, hanem annak tesztje, hogy az alapmodell mennyire jól viseli a magyar nyelvű finomhangolást.
Az eredeti GitHub-repó jelenleg nem tartalmaz kész merge kódot a LoRA és az alapmodell egyesítéséhez, bár ezt a kodóló AI-k megoldották helyettem, azaz mergeltem a lorat az eredetii modellbe. Új nyelv implementálásánál szakmailag inkább a teljes finomhangolás lenne javasolt.
Ennek ellenére a modell már megszólal magyarul. A kiejtés természetesen még nem tökéletes, de kísérletezésre már használható, és érezhetően jobban generál magyar beszédet, mint az alapmodell.
Miért alpha?
Ezt a verziót nem bétának, hanem inkább alpha állapotúnak tekintem.
A modell működőképes, de még nem végleges, nem stabil, és nem tekinthető kész magyar TTS modellnek.
Felhasználási korlátozások
A modell nem használható megszemélyesítésre, csalásra, megtévesztésre vagy dezinformációra.
AI-generált beszéd közzétételekor javasolt egyértelműen jelezni, hogy a hang szintetikus.
A kísérleti magyar finomhangolás a Common Voice 17.0 magyar adatrészén készült.
Dataset: fsicoli/common_voice_17_0
Dataset licence: CC0-1.0 / Public Domain
VoxCPM2
VoxCPM2 is a tokenizer-free, diffusion autoregressive Text-to-Speech model — 2B parameters, 30 languages, 48kHz audio output, trained on over 2 million hours of multilingual speech data.
Highlights
- 🌍 31-Language Multilingual — No language tag needed; input text in any supported language directly
- 🎨 Voice Design — Generate a novel voice from a natural-language description alone (gender, age, tone, emotion, pace…); no reference audio required
- 🎛️ Controllable Cloning — Clone any voice from a short clip, with optional style guidance to steer emotion, pace, and expression while preserving timbre
- 🎙️ Ultimate Cloning — Provide reference audio + its transcript for audio-continuation cloning; every vocal nuance faithfully reproduced
- 🔊 48kHz Studio-Quality Output — Accepts 16kHz reference; outputs 48kHz via AudioVAE V2's built-in super-resolution, no external upsampler needed
- 🧠 Context-Aware Synthesis — Automatically infers appropriate prosody and expressiveness from text content
- ⚡ Real-Time Streaming — RTF as low as ~0.3 on NVIDIA RTX 4090, and ~0.13 accelerated by Nano-VLLM
- 📜 Fully Open-Source & Commercial-Ready — Apache-2.0 license, free for commercial use
Supported Languages (30)
Arabic, Burmese, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, Vietnamese
Chinese Dialects: 四川话, 粤语, 吴语, 东北话, 河南话, 陕西话, 山东话, 天津话, 闽南话
Quick Start
Installation
Requirements: Python ≥ 3.10, PyTorch ≥ 2.5.0, CUDA ≥ 12.0 ·
Full Quick Start →
Text-to-Speech
1from voxcpm import VoxCPM
2import soundfile as sf
3
4model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
5
6wav = model.generate(
7 text="VoxCPM2 brings multilingual support, creative voice design, and controllable voice cloning.",
8 cfg_value=2.0,
9 inference_timesteps=10,
10)
11sf.write("output.wav", wav, model.tts_model.sample_rate)
Voice Design
Put the voice description in parentheses at the start of text, followed by the content to synthesize:
1wav = model.generate(
2 text="(A young woman, gentle and sweet voice)Hello, welcome to VoxCPM2!",
3 cfg_value=2.0,
4 inference_timesteps=10,
5)
6sf.write("voice_design.wav", wav, model.tts_model.sample_rate)
Controllable Voice Cloning
1# Basic cloning
2wav = model.generate(
3 text="This is a cloned voice generated by VoxCPM2.",
4 reference_wav_path="speaker.wav",
5)
6sf.write("clone.wav", wav, model.tts_model.sample_rate)
7
8# Cloning with style control
9wav = model.generate(
10 text="(slightly faster, cheerful tone)This is a cloned voice with style control.",
11 reference_wav_path="speaker.wav",
12 cfg_value=2.0,
13 inference_timesteps=10,
14)
15sf.write("controllable_clone.wav", wav, model.tts_model.sample_rate)
Ultimate Cloning
Provide both the reference audio and its exact transcript for maximum fidelity. Pass the same clip to both reference_wav_path and prompt_wav_path for highest similarity:
1wav = model.generate(
2 text="This is an ultimate cloning demonstration using VoxCPM2.",
3 prompt_wav_path="speaker_reference.wav",
4 prompt_text="The transcript of the reference audio.",
5 reference_wav_path="speaker_reference.wav",
6)
7sf.write("hifi_clone.wav", wav, model.tts_model.sample_rate)
Streaming
1import numpy as np
2
3chunks = []
4for chunk in model.generate_streaming(text="Streaming is easy with VoxCPM!"):
5 chunks.append(chunk)
6wav = np.concatenate(chunks)
7sf.write("streaming.wav", wav, model.tts_model.sample_rate)
Model Details
| Property | Value |
|---|
| Architecture | Tokenizer-free Diffusion Autoregressive (LocEnc → TSLM → RALM → LocDiT) |
| Backbone | Based on MiniCPM-4, totally 2B parameters |
| Audio VAE | AudioVAE V2 (asymmetric encode/decode, 16kHz in → 48kHz out) |
| Training Data | 2M+ hours multilingual speech |
| LM Token Rate | 6.25 Hz |
| Max Sequence Length | 8192 tokens |
| dtype | bfloat16 |
| VRAM | ~8 GB |
| RTF (RTX 4090) | ~0.30 (standard) / ~0.13 (Nano-vLLM) |
Performance
VoxCPM2 achieves state-of-the-art or competitive results on major zero-shot and controllable TTS benchmarks.
See the
GitHub repo for full benchmark tables (Seed-TTS-eval, CV3-eval, InstructTTSEval, MiniMax Multilingual Test).
Fine-tuning
VoxCPM2 supports both full SFT and LoRA fine-tuning with as little as 5–10 minutes of audio:
1# LoRA fine-tuning (recommended)
2python scripts/train_voxcpm_finetune.py \
3 --config_path conf/voxcpm_v2/voxcpm_finetune_lora.yaml
4
5# Full fine-tuning
6python scripts/train_voxcpm_finetune.py \
7 --config_path conf/voxcpm_v2/voxcpm_finetune_all.yaml
See the
Fine-tuning Guide for full instructions.
Limitations
- Voice Design and Style Control results may vary between runs; generating 1–3 times is recommended to obtain the desired output.
- Performance varies across languages depending on training data availability.
- Occasional instability may occur with very long or highly expressive inputs.
- Strictly forbidden to use for impersonation, fraud, or disinformation. AI-generated content should be clearly labeled.
Citation
1@article{voxcpm2_2026,
2 title = {VoxCPM2: Tokenizer-Free TTS for Multilingual Speech Generation, Creative Voice Design, and True-to-Life Cloning},
3 author = {VoxCPM Team},
4 journal = {GitHub},
5 year = {2026},
6}
7
8@article{voxcpm2025,
9 title = {VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning},
10 author = {Zhou, Yixuan and Zeng, Guoyang and Liu, Xin and Li, Xiang and
11 Yu, Renjie and Wang, Ziyang and Ye, Runchuan and Sun, Weiyue and
12 Gui, Jiancheng and Li, Kehan and Wu, Zhiyong and Liu, Zhiyuan},
13 journal = {arXiv preprint arXiv:2509.24650},
14 year = {2025},
15}
License
Released under the
Apache-2.0 license, free for commercial use. For production deployments, we recommend thorough testing and safety evaluation tailored to your use case.
