FarmerlineML/voxcpm2-swahili-sft

VoxCPM2 Swahili — Full SFT

A full fine-tune of VoxCPM2 on Swahili speech data. The model was trained by FarmerlineML for the Darli AI agricultural voice assistant.

The model generates Swahili speech using a reference recording to preserve the target speaker's voice characteristics.

Training Details

Parameter	Value
Base model	openbmb/VoxCPM2 (2B)
Language	Swahili / Kiswahili (sw)
Method	Full SFT—all model parameters trained
Learning rate	1e-5
Batch size	1
Gradient accumulation	16
Effective batch size	16
AudioVAE input sample rate	16 kHz
Final training step	2985

Validation Loss

Step	loss/total	loss/diff	loss/stop
0	1.301021	0.976434	0.216392
500	0.920041	0.854645	0.043597
1000	0.904256	0.847106	0.038100
1500	0.967527	0.931794	0.023822
2000	0.899414	0.846611	0.035202
2500	0.998210	0.927028	0.047455

Dataset

• FarmerlineML/SwahiliTTS2025_dataset

Update the dataset repository ID in this model card if your Swahili training dataset uses a different Hugging Face repository.

Usage

from voxcpm import VoxCPM
import numpy as np
import soundfile as sf

model = VoxCPM.from_pretrained(
    "FarmerlineML/voxcpm2-swahili-sft",
    load_denoiser=False,
)

def trim_audio(
    wav,
    sr,
    silence_thresh=0.01,
    max_silence_secs=2.0,
):
    """Trim a long silent tail from generated audio."""
    abs_wav = np.abs(wav)
    window = int(0.05 * sr)
    n_windows = len(abs_wav) // window
    max_silent_windows = int(max_silence_secs / 0.05)

    silence_count = 0
    cut_sample = len(wav)

    for window_index in range(n_windows):
        start = window_index * window
        end = (window_index + 1) * window
        chunk = abs_wav[start:end]

        if chunk.max() < silence_thresh:
            silence_count += 1

            if silence_count >= max_silent_windows:
                cut_sample = (
                    window_index - max_silent_windows + 1
                ) * window
                break
        else:
            silence_count = 0

    padding = int(0.1 * sr)
    final_sample = min(cut_sample + padding, len(wav))

    return wav[:final_sample]

text = "Habari za asubuhi. Karibu kwenye huduma ya Farmerline."

wav = model.generate(
    text=text,
    reference_wav_path="your_swahili_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)

The example sentence means: “Good morning. Welcome to the Farmerline service.”

Repository Structure

├── model.safetensors          # Fine-tuned model weights
├── audiovae.pth               # AudioVAE decoder
├── config.json                # Model architecture configuration
├── tokenizer.json             # Tokenizer vocabulary and rules
├── tokenizer_config.json      # Tokenizer configuration
├── special_tokens_map.json    # Special-token definitions
├── training/
│   ├── train.log              # Complete training log
│   ├── val_loss_summary.txt   # Validation losses by checkpoint
│   ├── training_state.json    # Final training state
│   └── scheduler.pth          # Learning-rate scheduler state
└── tensorboard/               # TensorBoard event files

Notes

• A reference audio recording is required during inference to anchor the generated audio to a speaker identity.
• The reference recording should contain clean Swahili speech with minimal background noise and reverberation.
• Use max_len=max(50, len(text) * 4) to reduce the risk of hallucinated audio after the sentence ends.
• Trimming long silent tails after generation is recommended.
• Generated output in this example is saved at 48 kHz.