VoxCPM2 Arabic: Saudi LoRA training checkpoint
The training checkpoint of the Saudi dialect LoRA for VoxCPM2, with optimizer and scheduler state so training can
be resumed. For speech synthesis you do not need this repo: the same LoRA weights are the saudi adapter in
VoxCPM2-Arabic-Dialects.
At a glance
| Base model | openbmb/VoxCPM2 |
| Dialect | Saudi |
| Training step | 78,000 |
| LoRA | rank 64, alpha 128, dropout 0 |
| LoRA targets | q_proj, k_proj, v_proj, o_proj in the LM and in the LocDiT; projections not adapted |
| Same weights as | adapters/saudi/lora_weights.safetensors in VoxCPM2-Arabic-Dialects (SHA-256 identical) |
| Visibility | Private |
Files
| File | Size | What it is |
|---|---|---|
lora_weights.safetensors |
144.7 MB | LoRA weights |
lora_config.json |
0.5 kB | LoRA settings (table above) |
optimizer.pth |
289.7 MB | Optimizer state, for resuming training |
scheduler.pth |
1.5 kB | Learning-rate scheduler state |
training_state.json |
โ | {"step": 78000} |
lora_weights.safetensors and lora_config.json have the same layout as each adapters/<dialect>/ folder that the
engine loads.
Usage
Inference (recommended), with voxcpm2-arabic-dialects-turbo:
tts = await DialectTTS.load("VoxCPM-Kernel/VoxCPM2-Arabic-Dialects")
await tts.add_voice("ahmed", "ahmed.wav")
audio = await tts.synthesize("ููุง ูุงูููุ ููู ุญุงูู ุงูููู
ุ", dialect="saudi", voice="ahmed")
Source
Copied from mnawahda/voxcpm_saudi_22. No licence was supplied with it; the base model VoxCPM2 is Apache-2.0.
Model tree for VoxCPM-Kernel/VoxCPM2-Arabic-Saudi-LoRA-Checkpoint
Base model
openbmb/VoxCPM2