Sinsy

Sinsy (Singing Voice Synthesis System) is where open singing synthesis started. Built by Keiichi Tokuda's group at the Nagoya Institute of Technology under a Modified BSD license, it began as an HMM-based statistical parametric system and moved to deep neural networks in 2016. The DNN version is the subject of the 2021 IEEE/ACM TASLP paper. The public web demo is genuinely usable: you upload a MusicXML score with lyrics and the site renders up to five minutes of sung audio, with controls for gender factor, vibrato intensity and pitch shift. The catch, documented in the NNSVS paper: the publicly released open-source code only covers the older HMM-based system. The DNN version behind the website was never published. Sinsy's modular pipeline (time-lag, duration, acoustic model, vocoder) became the template that NNSVS and others copied, and the DNN paper reported pitch correction and vibrato modeling that made it competitive with commercial engines of its era (author-reported listening tests).

• DNN version reported competitive naturalness vs commercial SVS engines in the TASLP 2021 paper's listening tests (author-reported)
• Pitch-correction technique shown to suppress out-of-tune artifacts in training data (author-reported)
• No published benchmark figures for the web demo; no independent replication found