HiFiSinger
Microsoft's answer to why singing synthesis sounded lo-fi: a 48 kHz system with a FastSpeech acoustic model, a Parallel WaveGAN vocoder, and GAN training split across frequency sub-bands.
About
HiFiSinger (arXiv 2009.01776, September 2020) tackles a concrete problem: sung voices carry expression in frequency ranges and sustained vowels that 16 to 24 kHz speech systems smear away. The Microsoft Research system runs at 48 kHz with a FastSpeech-based neural acoustic model and a Parallel WaveGAN-based vocoder, plus three ideas the authors call crucial: a sub-frequency GAN that trains separate discriminators for low/mid/high mel bands, a multi-length GAN for the longer waveforms 48 kHz produces, and F0 plus voiced/unvoiced flags as explicit acoustic features. The paper reports a 0.32 to 0.44 MOS gain over 48/24 kHz baselines and 0.83 MOS over previous SVS systems (author-reported), with demo audio still hosted on the project page. Like DeepSinger, Microsoft never published official code or weights; the GitHub repos under the name are unofficial reimplementations. It sits in Microsoft's Muzic project lineage alongside XiaoiceSing.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- +0.32/0.44 MOS over 48kHz/24kHz baselines, +0.83 MOS over previous SVS systems (author-reported)
- Ablations for the sub-frequency GAN, multi-length GAN, pitch input, and window/hop choices (author-reported)
- No official code/weights, so no independent reproduction exists
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER verified, metrics-v1). Evidence: Microsoft Research paper with a real architectural contribution (sub-frequency and multi-length GAN training for 48 kHz singing) and thorough ablations. Strengths: one of the first serious attempts at high-fidelity SVS, and its design notes (F0 as an input feature, receptive-field choices) are still cited as practical guidance. Docked for: (1) no official code or weights were ever released, so it is not reproducible or usable; (2) every MOS figure is author-reported; (3) dormant since 2020. License: no 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
- 2.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · arXiv 2009.01776, September 2020; part of Microsoft's Muzic project
- · 48 kHz: sub-frequency GAN + multi-length GAN, FastSpeech + Parallel WaveGAN
- · Reported +0.83 MOS over previous SVS systems (author-reported)
- · No official code or weights ever released; dormant since 2020