Nepali RVC Voices v1
Four independently trained Nepali RVC target voices. This is a voice pack,
not a blended checkpoint: select one .pth generator and its matching
.index file at inference time. Averaging the four checkpoints would erase
speaker identity and make each retrieval index inconsistent with its model.
The canonical machine-readable inventory is voices.json.
All release files are SHA-256 pinned there.
Voices
| ID | Presentation | Clean clips | Duration |
|---|---|---|---|
ne_rvc_f01 | female | 165 | 27.572 min |
ne_rvc_f02 | female | 170 | 26.798 min |
ne_rvc_m01 | male | 156 | 23.721 min |
ne_rvc_m02 | male | 155 | 21.254 min |
Speakers remain identified only by dataset pseudonymous IDs. These voices must not be presented as a real named person, used to deceive listeners, or used without rights to the source audio being converted.
Training and inference
- Runtime: Applio RVC
- 48 kHz, 80 epochs, HiFi-GAN
- RMVPE F0 extraction and ContentVec features
- Default retrieval index rate:
0.6 - Default pitch:
0semitones - Dataset audio transformations: stereo-to-mono mean and PCM16 encoding only
- No pitch shift, phase transformation, denoising, or pseudovoice processing
Browser-native exports
The browser/ directory contains fixed 512-frame FP16 ONNX
generators and compact retrieval graphs for every voice. The compact graph
uses the learned IVF coarse centroids from the corresponding FAISS index,
reducing each browser index to 5–7 MB while preserving local retrieval.
The static application at
voidash/nepali-rvc-tts
runs the Nepali frontend, carrier TTS, ContentVec, RMVPE, retrieval, and RVC
synthesis on the client device. Its host serves static files only; source text
and audio are not submitted to a model API.
Attribution and license
Training audio is from
ai4bharat/indicvoices_r,
released under CC BY 4.0. The voice pack is released under CC BY 4.0 and must
retain this attribution. RVC inference is provided through
IAHispano/Applio, whose code is MIT
licensed and whose terms of use also apply.
This is an experimental release. Listen and evaluate suitability before any production deployment.