Skip to content
← Models
Open weightsHaojie Wei et al. (Renmin University of China)/10 trustVerified

RMVPE

The pitch tracker the RVC voice-conversion community standardized on: a deep U-Net plus GRU that pulls vocal F0 straight out of polyphonic music, no source separation needed.

Visit official page →

About

RMVPE (arXiv 2306.15412, June 2023; Interspeech 2023) is a vocal pitch estimation model built for the hard case: extracting a singer's fundamental frequency from a full mix with drums, bass and guitars still in it. The network takes a 16 kHz log-mel spectrogram through a deep U-Net with BiGRU layers and predicts a 360-bin pitch-probability matrix at 10 ms per frame, decoded with a local-argmax scheme. The authors report superiority over prior vocal-pitch estimators in raw pitch accuracy (RPA) and raw chroma accuracy (RCA), plus robustness across SNR levels, though those numbers are author-reported. The official repo (Dream-High/RMVPE, Apache 2.0) ships code only. The weights most people actually use come from the RVC community mirror (lj1995/VoiceConversionWebUI, tagged MIT). RMVPE became the recommended default pitch method in RVC and its forks, which is why it matters well beyond its citation count.

Specs

Context window

16 kHz mono input; frames padded to multiples of 32, output at 100 fps; no published max-duration limit.

Pricing

Free / open weights

Free quota

Fully open: code on GitHub (Apache 2.0); weights via the MIT-tagged RVC community mirror

API access

No hosted API: run locally via Python, ONNX, or the RVC WebUI

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Raw pitch accuracy (RPA) and raw chroma accuracy (RCA) improvements over prior estimators on polyphonic music (author-reported)
  • Robustness verified by the authors across noise types and SNR levels (author-reported)
  • No independent benchmark replication found; adoption (default pitch method in RVC) is the practical evidence

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

vocal-pitch-estimationsinging-voice-conversionmusic-transcriptionautotunekaraoke-scoring

Trust score — 7/10

/10 trustVerified

Scored 7/10 (TIER verified, metrics-v1). Evidence: peer-reviewed at Interspeech 2023; Apache 2.0 official code; the RVC ecosystem adopted it as the default pitch extractor, which is stronger real-world validation than most citations. Strengths: it solves vocal pitch estimation on polyphonic mixes without a source-separation step, and the community has battle-tested it on messy real audio. Docked for: (1) the official repo ships no weights, so the redistributable weight path is a community mirror with secondhand provenance; (2) every accuracy figure is author-reported with no independent replication; (3) the official repo has been quiet since 2023. License: Apache 2.0 (code); community weights tagged MIT.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
4.0/10
Openness · license + docs
7.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · Peer-reviewed: Interspeech 2023, arXiv 2306.15412
  • · Official repo ships code only; weights provenance is secondhand
  • · Default pitch extractor in RVC ecosystem
  • · Official repo quiet since 2023

Sources