You can deploy this model and start earning money today!

fahmifauzi/
rvc-malaysian-voices

Name: rvc-malaysian-voices
Availability: InStock

mesolitica/tts-combine-annotated

mit

Views

No views yet

RVC Malaysian Voice Models

Pre-trained RVC v2 voice conversion models for Malaysian Malay voices. Designed for integration with TTS pipelines to produce natural-sounding Malaysian speech.

Models

Model	Voice	Gender	Training Data	FAISS Vectors
`suara_wanita_1.pth`	Yasmin	Female	22.7 min / 317 segments	67,907
`suara_wanita_2.pth`	Bunga	Female	37.9 min / 454 segments	113,281
`suara_lelaki_1.pth`	Osman	Male	21.8 min / 309 segments	65,313
`suara_lelaki_2.pth`	Ariff	Male	40.3 min / 464 segments	120,560

Training Details

Architecture: RVC v2, 40kHz sample rate
Epochs: 300 per model
F0 Method: RMVPE
Features: HuBERT (256-dim to 768-dim)
Pretrained weights: f0G40k.pth / f0D40k.pth
Infrastructure: Modal.com (L4 GPU)

Usage

With `infer-rvc-python`

python

from infer_rvc_python import BaseLoader
from huggingface_hub import hf_hub_download

# Download model and index
pth = hf_hub_download("fahmifauzi/rvc-malaysian-voices", "suara_wanita_1.pth")
idx = hf_hub_download("fahmifauzi/rvc-malaysian-voices", "suara_wanita_1.index")

# Initialize converter
converter = BaseLoader(only_cpu=False)
converter.apply_conf(
    tag="convert",
    file_model=pth,
    pitch_algo="rmvpe+",
    pitch_lvl=0,
    file_index=idx,
    index_influence=0.75,
    envelope_ratio=0.25,
    consonant_breath_protection=0.5,
)

# Convert audio
result = converter(["input.wav"], "convert")
print(f"Output: {result[0]}")

Loading a specific checkpoint

Intermediate checkpoints are available under checkpoints/ for A/B comparison:

python

pth = hf_hub_download(
    "fahmifauzi/rvc-malaysian-voices",
    "checkpoints/suara_wanita_1/suara_wanita_1_e200_s8400.pth",
)

Recommended Parameters

Parameter	Value	Description
`pitch_algo`	`rmvpe+`	Pitch extraction algorithm
`pitch_lvl`	`0`	Pitch shift in semitones (-12 to +12)
`index_influence`	`0.75`	FAISS index retrieval weight
`envelope_ratio`	`0.25`	Volume envelope mix ratio
`consonant_breath_protection`	`0.5`	Protects consonants from artifacts

Repo Structure

├── suara_wanita_1.pth          # Final model weights
├── suara_wanita_1.index        # FAISS retrieval index
├── suara_wanita_2.pth
├── suara_wanita_2.index
├── suara_lelaki_1.pth
├── suara_lelaki_1.index
├── suara_lelaki_2.pth
├── suara_lelaki_2.index
└── checkpoints/                # Intermediate checkpoints (every 50 epochs)
    ├── suara_wanita_1/
    ├── suara_wanita_2/
    ├── suara_lelaki_1/
    └── suara_lelaki_2/

Acknowledgements

Training data sourced from datasets by Mesolitica:

mesolitica/tts-combine-annotated — Synthetic Malaysian TTS dataset with quality annotations (SNR, STOI, PESQ)

Quality filters applied during extraction: SNR >= 20 dB, STOI >= 0.85, PESQ >= 3.0, duration 2-30s.

fahmifauzi/rvc-malaysian-voices