VoxCPM2 — prebuilt weight sets for audio.cpp
VoxCPM2 converted for audio.cpp, cut exactly the way the Voicebox-on-Linux panel splits the model: the language part (BaseLM) and the acoustic part (ResidualLM, LocEnc, DiT, projections) each in its own precision. AudioVAE (the head) is always F32.
| set | BaseLM | acoustic | size |
|---|---|---|---|
q8_0-f16/ |
Q8_0 | F16 | 3.6 GB |
q8_0-f32/ |
Q8_0 | F32 | 5.2 GB |
f16-f16/ |
F16 | F16 | 5.0 GB |
f16-f32/ |
F16 | F32 | 6.6 GB |
Why prebuilt: without them the engine takes the BF16 original and converts it into the chosen precision on every load — slow (~7 s, ~23 s from a cold disk) and RAM can spike hard while both copies sit in memory. A prebuilt set loads in ~2 s.
The tensors are bit for bit what the engine would produce on the fly (same ggml conversion):
generator biases and norms F32, BaseLM matrices → BaseLM type, the rest of the generator →
acoustic type, AudioVAE untouched. Made with audiocpp/make_weights.py from the repo above —
run it yourself if you prefer.
Use
Each set is a directory with voxcpm2.gguf, laid out as Voicebox-on-Linux expects:
models/audiocpp/prebuilt/<set>/voxcpm2.gguf (./download-models.sh all fetches all four).
Tested with the patched audio.cpp from that repo (audiocpp/voicebox.patch, @ 497c7f0).
License
Apache 2.0, as the original openbmb/VoxCPM2.
- Downloads last month
- 569
16-bit
32-bit
Model tree for grruwi/VoxCPM2-audiocpp-prebuilt
Base model
openbmb/VoxCPM2