RVC Voice Conversion and AI Song Covers · TelkNet

Voice & singingGPU workloadAI-RVC v1.6.0 · Changelog RVC · Source AI-RVC · Source RVC v2 · Paper RMVPE · Source RMVPE · Paper MelBandRoformers · Weights bs-roformer-karaoke · Weights karaoke_bs_roformer · Weights MVSep 9205 BS-Kar AVG (lead/back vocals) · Paper BS-Roformer-Leap · Weights python-audio-separator · Source demucs · Source Leap XE / Leap Instrumental / Karaoke separation · Paper Leap XE / Leap Instrumental / Karaoke separation · Paper

AI Cover Vocals

30 credits/useAudio file inputMax 2048 MBEstimated 180-900 seconds

Input

Clean vocals, cover material, or separated vocal stems through an RVC v2 conversion chain

Audio formats

mp3wavflacoggm4aaac

Output

Converted vocals, original vocal references, and downloadable audio for demos and listening tests

Use case

Voice-color conversion, cover demos, and character voice experiments

Quick feedback

Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.

Models, parameters, and sources

The AI cover pipeline combines separation, RVC voice conversion, and mixdown. The core inputs are a vocal file and target voice model.

RVC v2RMVPEMVSep 9205 BS-Kar AVG (lead/back vocals)BS-RoFormer Leap XE 90 bands (vocals pre-separation)BS-RoFormer Leap Instrumental (accompaniment pre-separation)Leap XE / Leap Instrumental / Karaoke separationDemucs / UVR5 fallbackRMVPE Interspeech Table 1 — MIR-1KRMVPEPolyCREPEHARMOF0JDC

RVC v2

RVC v2 converts the input vocal timbre into the selected character voice model.

RVC Project TelkNet AI-RVC workflow AI-RVC v1.6.0 HuBERT

Standard vocal/accompaniment separation. BS-RoFormer Leap XE 90 bands outputs the vocal stem, and BS-RoFormer Leap Instrumental outputs the accompaniment stem.
Voice model: selects the target character weights and retrieval index.

RMVPE

AI-RVC v1.6.0 uses RMVPE to extract vocal pitch by default.

RVC Project Dream-High/RMVPE RMVPE

Pitch role: tracks source vocal F0 before conversion so the target timbre follows the melody.
This mode uses RMVPE to extract pitch.
Public evidence: the RMVPE Interspeech paper reports top RPA/RCA/OA on multiple polyphonic vocal pitch datasets.

MVSep 9205 BS-Kar AVG (lead/back vocals)

MVSep 9205 supplies the lead/accompaniment-with-harmony leg, BS-RoFormer Leap XE 90 bands supplies vocals with harmony, and BS-RoFormer Leap Instrumental supplies the accompaniment reference.

MVSep 9205 GaboxR67 bs_karaoke_gabox_IS becruily Frazer/Becruily weights anvuew karaoke_bs_roformer Mel-Band RoFormer

Active MVSep 9205 checkpoints: BS-Kar-Gabox_IS, BS-Kar-Frazer&Becruily, and BS-Kar-Anvuew with avg_wave ensemble.
MVSep 9205 outputs: vocals_without_harmony plus backing+instrumental from the original full mix.

Leap XE / Leap Instrumental / Karaoke separation

The AI cover pipeline separates vocals/accompaniment before conversion and mixdown.

pcunwa Leap XE weights MVSep 10178 Leap XE Leap Instrumental configuration Leap Instrumental checkpoint python-audio-separator facebookresearch/demucs MVSep algorithms BS-RoFormer HT Demucs

Primary separator: Leap XE 90 prepares vocals and Leap Instrumental prepares accompaniment before RVC.
Harmony path: MVSep 9205 lead/back can export lead and backing vocal stems before RVC mixing.
Separator: uses RoFormer, UVR5, or Demucs before conversion.

Model comparison

Model

Role

Why it is used

Output impact

ModelRVC v2

RVC Project TelkNet AI-RVC workflow AI-RVC v1.6.0 HuBERT

RoleVoice conversion core

Why it is usedRVC v2 supplies the selected target timbre and index-assisted conversion.

Output impactConverted vocals

ModelRMVPE

RVC Project Dream-High/RMVPE RMVPE

RolePitch extraction

Why it is usedPitch role: tracks source vocal F0 before conversion so the target timbre follows the melody.

Output impactMelody/F0 contour

ModelBS-RoFormer Leap XE 90 bands (vocals pre-separation)

pcunwa Leap XE weights MVSep 10178 Leap XE BS-RoFormer

RoleVocal preparation

Why it is usedStandard vocal/accompaniment separation. BS-RoFormer Leap XE 90 bands outputs the vocal stem, and BS-RoFormer Leap Instrumental outputs the accompaniment stem.

Output impactVocals and accompaniment stems

ModelBS-RoFormer Leap Instrumental (accompaniment pre-separation)

Leap Instrumental configuration Leap Instrumental checkpoint BS-RoFormer

RoleVocal preparation

Why it is usedStandard vocal/accompaniment separation. BS-RoFormer Leap XE 90 bands outputs the vocal stem, and BS-RoFormer Leap Instrumental outputs the accompaniment stem.

Output impactVocals and accompaniment stems

ModelDemucs / UVR5 fallback

facebookresearch/demucs HT Demucs python-audio-separator

RoleCompatibility separator

Why it is usedDemucs/UVR5 remain useful when a specific source behaves better outside RoFormer.

Output impactAlternative stems

All evaluation sources (1)

Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.

RMVPE · Table 1 ↗
Deployable entries 0 / source rows 6
Not ranked: 6
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
PYIN [6] · CREPE [7] · HARMOF0 [9] · CRN-Raw [13] · JDC [14] · RMVPEPoly

Paper and benchmark notes

HuBERT: Self-Supervised Speech Representation Learning

IEEE/ACM TASLP - 2021

HuBERT provides self-supervised speech representations used by RVC-style voice conversion pipelines.

arXiv:2106.07447

RMVPE: A Robust Model for Vocal Pitch Estimation in Polyphonic Music

Interspeech - 2023

RMVPE is an Interspeech 2023 robust vocal pitch estimator designed for polyphonic music and used for quality-first F0 extraction.

ISCA PDF Dream-High/RMVPE

RVC Project

Official repository - current

The RVC project is the official technical source for the VITS/HuBERT retrieval-based voice conversion workflow.

RVC WebUI AI-RVC workflow AI-RVC v1.6.0

Parameter guide

Voice model: selects the target character weights and retrieval index.
Pitch shift: adjusts the source by semitones to match the target range.
Index ratio: controls how much the retrieval index affects voice color.
Official Mode · Use RVC/Applio UVR5 separation and voice conversion, without additional dereverb or post-processing.
Mix controls: adjust vocals, accompaniment, reverb, and original-vocal blend.

Use cases

Build cover demos and character-voice trials
Validate a voice direction before deeper production
Feed separated vocals into a downstream singing workflow

Workflow

1Prefer to upload already-separated vocal material
2Submit the conversion task with the target voice direction
3Listen to the result before deciding on further mixing or polishing

What to check

The cleaner the vocal, the more stable the conversion result
Separating vocals first usually beats feeding mixed audio directly into RVC
Check consonants, sustained notes and high notes for clear pronunciation, pitch jumps and vocal level when mixed with the accompaniment.

Related use-case guides

Find processing methods and steps suited to your material.

RVC Voice Conversion · online tool and model guide AI cover creation · online tool and model guide ZONOS2 Voice Cloning TTS · online tool and model guide