DeepSinger
The first singing synthesizer trained entirely on data mined from the web: 92 hours, 89 singers, three languages. No studio recordings, no manual alignment.
About
DeepSinger (KDD 2020; arXiv 2007.04590) asked whether singing synthesis could skip the studio entirely. The Microsoft Research team built a full pipeline: crawling music sites, separating vocals from accompaniment, aligning lyrics to singing with a coarse-to-fine phoneme alignment model, filtering the results, then training a multilingual multi-singer singing model on a feed-forward Transformer that generates linear spectrograms from lyrics (synthesized with Griffin-Lim). The mined dataset was about 92 hours from 89 singers across Chinese, Cantonese and English, with a reference encoder capturing each singer's timbre from the noisy web audio. The paper reports high-quality output in pitch accuracy and voice naturalness from purely web-mined data, and demonstrated cross-lingual singing as a bonus (all author-reported). The honest limitation: Microsoft never released official code or weights. What exists on GitHub are incomplete community reimplementations. The demo page with audio samples is still up.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper reports high pitch accuracy and naturalness from web-mined data alone, in subjective listening tests (author-reported)
- Demonstrated cross-lingual synthesis as a byproduct of multilingual training (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 KDD 2020 with a genuinely novel contribution, web-scale data mining for SVS plus a lyrics-to-singing alignment model that removed manual labeling. Strengths: the mined-data pipeline idea influenced the whole subfield, and the 92-hour multilingual dataset construction was unprecedented. Docked for: (1) Microsoft never released code or weights, so nothing about DeepSinger is reproducible or usable today; (2) all quality claims are author-reported; (3) the project is 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
- · Peer-reviewed: KDD 2020, arXiv 2007.04590
- · 92 hours / 89 singers / 3 languages mined from the web
- · No official code or weights ever released
- · Dormant since 2020; demo page still live