Evidence and limitations · Updated August 7, 2026

How NiceVois verifies its RVC training workflow

This page separates implemented behavior, observed test results, practical product recommendations, and facts taken from the upstream RVC Project.

Pipeline under test

Validate and privately store supported source audio.
Slice and resample it for the pinned 48 kHz RVC v2 training path.
Extract RMVPE pitch and HuBERT v2 voice features.
Fine-tune the generator and discriminator from pinned pretrained assets.
Report completed epochs and publish validated inference checkpoints at configured intervals.
Export and load the final .pth through the native RVC synthesizer.
Build and validate the FAISS .index, then package both files into a ZIP.
Verify file size and SHA-256 when copying completed artifacts to private backup storage.

Observed end-to-end results

Observation	Verified result
Input used for the first hosted proof	33,189,998-byte WAV
20-epoch hosted run	Browser upload through named .pth download completed; resulting model was 57,598,812 bytes
Elapsed job time in the persisted record	Approximately 13 minutes 35 seconds from creation to completion
100-epoch RVC v2 proof	53 prepared segments; 57,590,877-byte .pth; native RVC load passed
Portability checks	Required config, pitch, sample-rate, version, and weight fields were present

What these tests prove: the specific inputs completed the implemented upload, GPU, export, validation, and download path. They do not prove that every voice, recording, provider queue, or epoch count will take the same time or sound equally good.

How recommendations are formed

The epoch dial uses a conservative duration rule: 25 epochs under one minute, 100 below five minutes, 150 below ten minutes, and 200 for ten minutes or more. The public beta caps runs at five minutes and 100 epochs. This is a product starting point, not a universal RVC law. Recording cleanliness, speaker consistency, expression, and the intended conversion material can matter more than duration alone.

Known sources of variation

Lossy compression, music bleed, reverb, clipping, other speakers, and denoising artifacts.
Dataset duration and the amount of useful vocal variation.
Cloud provider queue time, GPU type, cold starts, and transient availability.
Epoch count and how closely the test material resembles the training performance.
Inference settings in the RVC interface used after download.

Primary technical references

Run the same product workflow

The workspace shows the actual stages, completed epochs, provider, and available artifacts for your job.

Open the training workspace