VoxCPM2 Krio — Full SFT
Full fine-tune of VoxCPM2 on Krio (Sierra Leone Creole) speech data. Trained by FarmerlineML for the darli AI agricultural voice assistant.
Training Details
| Parameter | Value |
|---|---|
| Base model | openbmb/VoxCPM2 (2B) |
| Method | Full SFT (all parameters) |
| Learning rate | 1e-5 |
| Batch size | 1 (grad accum 16, effective batch 16) |
| Sample rate | 16kHz (AudioVAE encoder input) |
| Final step | 1484 |
Validation Loss:
| Step | loss/total | loss/diff | loss/stop |
|---|---|---|---|
| 0 | 1.171295 | 0.762161 | 0.272756 |
| 200 | 0.780878 | 0.764130 | 0.011165 |
| 400 | 0.706954 | 0.694261 | 0.008462 |
| 600 | 0.693552 | 0.691856 | 0.001131 |
| 800 | 0.636625 | 0.634136 | 0.001659 |
| 1000 | 0.704854 | 0.676256 | 0.019065 |
| 1200 | 0.679028 | 0.677448 | 0.001054 |
| 1400 | 0.710708 | 0.652015 | 0.039129 |
Datasets:
Usage
from voxcpm import VoxCPM
import soundfile as sf
import numpy as np
model = VoxCPM.from_pretrained(
"FarmerlineML/voxcpm2-krio-sft",
load_denoiser=False,
)
def trim_audio(wav, sr, silence_thresh=0.01, max_silence_secs=2.0):
abs_wav = np.abs(wav)
window = int(0.05 * sr)
n_wins = len(abs_wav) // window
max_sil = int(max_silence_secs / 0.05)
silence_count, cut_sample = 0, len(wav)
for w in range(n_wins):
chunk = abs_wav[w * window:(w + 1) * window]
if chunk.max() < silence_thresh:
silence_count += 1
if silence_count >= max_sil:
cut_sample = (w - max_sil + 1) * window
break
else:
silence_count = 0
return wav[:min(cut_sample + int(0.1 * sr), len(wav))]
wav = model.generate(
text="Ow di bodi? Mi na Farmerline agent.",
reference_wav_path="your_krio_speaker.wav",
cfg_value=2.0,
inference_timesteps=15,
retry_badcase=False,
max_len=max(50, len(text) * 4),
)
wav = trim_audio(wav, 48000)
sf.write("output.wav", wav, 48000)
Repo Structure
├── model.safetensors # Model weights (~9.2GB)
├── audiovae.pth # AudioVAE decoder
├── config.json # Model architecture config
├── tokenizer.json # Tokenizer
├── training/
│ ├── train.log # Full training log
│ ├── val_loss_summary.txt # Validation losses per checkpoint
│ └── training_state.json # Final training state
└── tensorboard/ # TensorBoard event files
Notes
- Reference audio is required at inference for voice identity anchoring
- Use
max_len=max(50, len(text) * 4)to prevent hallucination after sentence end - A post-generation 2-second silence trim is strongly recommended
- Trained on Sierra Leone Krio — best results with Krio speaker reference audio