PeriodNet
A GAN vocoder that thinks in harmonics: two parallel generators turn a sine wave and noise into the periodic and aperiodic parts of a singing voice, staying natural even on pitches outside the training range.
About
PeriodNet (ICASSP 2021; arXiv 2102.07786) is a non-autoregressive neural vocoder designed specifically for the thing singing synthesis breaks other vocoders on: pitch. A waveform is really periodic and aperiodic components mixed together, and decomposing them in advance is error-prone, so PeriodNet feeds the two components' ideal inputs (a sine wave from F0, and Gaussian noise, plus voiced/unvoiced flags) into two parallel generators and lets the network learn the separation itself, summing the outputs. Experiments on a singing-voice corpus showed improved naturalness over baselines, and the headline result was more natural synthesis of pitches outside the training data's range (author-reported). The design was adopted as the pitch-robust vocoder inside Sinsy's DNN system, which is where most people encounter it. No official standalone code release could be verified; it survives through community reimplementations and as a component of larger systems.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Improved naturalness on a singing-voice corpus vs baselines (author-reported)
- More natural out-of-training-range pitch generation (author-reported)
- No independent replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: peer-reviewed at ICASSP 2021 with a clean, well-motivated idea, implicit periodic/aperiodic separation without error-prone preprocessing, and it was adopted as the vocoder inside Sinsy's DNN system, which is real-world validation. Strengths: the out-of-range pitch robustness directly addresses a known SVS failure mode. Docked for: (1) no official standalone code or weights could be verified; (2) all figures are author-reported; (3) it is a component model, not a usable end-to-end product. License: no standalone release, not applicable.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 3.0/10
- Openness · license + docs
- 5.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ICASSP 2021, arXiv 2102.07786; non-autoregressive GAN vocoder with periodic/aperiodic separation
- · Adopted as the pitch-robust vocoder inside Sinsy's DNN system
- · Component model, not an end-to-end product
- · Release date 2021-02-01 month-level approximation