XiaoiceSing
The FastSpeech-based singing synthesizer that made pitch stay on-key: residual F0 prediction and syllable-level duration constraints in one integrated network.
About
XiaoiceSing (Interspeech 2020; arXiv 2006.06261) is Microsoft's integrated singing voice synthesis system, named for the Xiaoice virtual singer project. It follows the FastSpeech architecture but adds singing-specific machinery: musical-score features (note pitch and length) alongside phoneme IDs, a residual connection between note pitch and predicted F0 to fight off-key drift, and a syllable-duration loss accumulated over all phonemes in a note to keep rhythm tight. The paper reports it beating a CNN baseline by 1.44 MOS on sound quality, 1.18 on pronunciation accuracy and 1.38 on naturalness, with A/B preference rates of 97.3% for the F0 modeling and 84.3% for duration modeling (all author-reported). Demo audio from the listening tests is still hosted. As with Microsoft's other SVS work of the era, no official code or weights were released, so it stands as a paper-only system that informed later designs like HiFiSinger.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- +1.44 MOS sound quality, +1.18 pronunciation, +1.38 naturalness over CNN baseline (author-reported)
- A/B tests: 97.3% preference for F0 modeling, 84.3% for duration modeling (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: peer-reviewed at Interspeech 2020 with clearly documented singing-specific innovations (residual F0, syllable-duration loss) and unusually detailed A/B testing. Strengths: the residual-F0 idea directly addressed the off-key problem and shows up in later systems. Docked for: (1) no official code or weights were ever released; (2) all figures are author-reported with no independent verification; (3) dormant since 2020. License: no release, not applicable.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 2.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · Interspeech 2020, arXiv 2006.06261; FastSpeech-based integrated SVS
- · Reported +1.44 MOS sound quality vs CNN baseline (author-reported); 97.3%/84.3% A/B preference for F0/duration (author-reported)
- · Residual F0 connection and syllable-duration loss are the key ideas
- · No official code or weights; dormant since 2020; release date 2020-06-01 month-level approximation