AEmotionStudio/voxcpm2-models

VoxCPM2 — inference mirror for MAESTRO

Inference-ready repack of openbmb/VoxCPM2 (OpenBMB), kept by AEmotionStudio so the MAESTRO DAW can fetch the checkpoint on demand. The model is not fine-tuned or otherwise altered — this repo only repacks the upstream weights for loading:

• model.safetensors — the 2B tokenizer-free diffusion-AR stack (LocEnc → MiniCPM-4 TSLM → RALM → LocDiT), verbatim from upstream (bf16 storage, sha256-checked; bf16 confirmed by MAESTRO's same-seed fp16-vs-bf16 whisper-gate against a seed-to-seed null baseline).
• audiovae.safetensors — AudioVAE V2 (16 kHz encode → 48 kHz decode), converted from the pickled audiovae.pth via safetensors.save_model (the VAE uses weight_norm, whose weight_g/weight_v aliasing a bare save_file would corrupt).
• config.json, tokenizer.json, special_tokens_map.json — verbatim.
• tokenizer_config.json — verbatim minus the list-form extra_special_tokens key (crashes some transformers loaders; the tokens are already registered in tokenizer.json, so dropping the key is lossless).
• upstream's tokenization_voxcpm2.py is omitted — inference loads a plain LlamaTokenizerFast; no trust_remote_code anywhere.
• sha256.json — checksums of every file in this repo.

Component	Purpose
model.safetensors	LocEnc + MiniCPM-4 2B TSLM + RALM + LocDiT
audiovae.safetensors	AudioVAE V2 (latents → 48 kHz mono waveform)
tokenizer files	LlamaTokenizerFast vocabulary + config

VoxCPM2 is a tokenizer-free diffusion-autoregressive TTS at native 48 kHz: 30 languages plus 9 Chinese dialects with no language tags, voice design from a text description, and three cloning tiers (reference clip / reference + style directive / reference + transcript).

License

• Code: Apache-2.0 (OpenBMB/VoxCPM).
• Weights (this repo): Apache-2.0 — inherited unchanged from upstream.

Credit for the model belongs entirely to the OpenBMB team; this mirror exists only for reliable on-demand delivery into MAESTRO.