Conv-TasNet
The fully-convolutional time-domain separator that beat ideal masking on speech — the foundation every modern separator builds on.
About
Conv-TasNet is the paper that changed what separation models are allowed to be. Before it, the ceiling was the 'ideal' time-frequency mask; Luo and Mesgarani showed a fully convolutional network operating directly on the waveform could surpass it on speech separation. The recipe — a learned encoder, a temporal convolutional network with dilated depthwise-separable convolutions, and a learned decoder, trained with scale-invariant SNR — became the template for DPRNN, SepFormer, and the whole time-domain family.\n\nThe headline numbers are speech-domain: 15.3 dB SI-SNR improvement on WSJ0-2mix (Asteroid's reproduction hits 16.2). But it earned its place in a music catalog honestly: the Sound Demixing Challenge 2023 leaderboard lists it as a music-demixing baseline at 5.73 average global SDR — far behind the specialists, which is itself informative about how far music separation has moved past generic separators.\n\nThe original reference implementation is open, and it ships in every toolkit (Asteroid, SpeechBrain, torchaudio). For music stems use a music model; for understanding where the field came from — or for speech/vocal-adjacent isolation — start here.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (IEEE/ACM TASLP 2019): 15.3 dB SI-SNRi / 15.6 dB SDRi on WSJ0-2mix; Asteroid reproduction 16.2/16.5 dB\n- Music context (SDX'23 leaderboard, ByteDance paper): 5.73 avg global SDR as a music-demixing baseline — well below music specialists\n- Speech-domain numbers; music performance is baseline-grade, not competitive
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: landmark peer-reviewed paper (IEEE/ACM TASLP 2019, arXiv 1809.07454) that surpassed ideal time-frequency masking, with reproduced numbers in Asteroid (16.2 dB SI-SNRi) and implementations in every major audio toolkit. Strengths: foundational, peer-reviewed, ubiquitously implemented, reproducible. Docked for: (1) speech-domain — music performance is baseline-grade; (2) 2019-era, superseded for any production use. License: open reference implementation; Asteroid/SpeechBrain ports are MIT/Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Landmark peer-reviewed paper, IEEE/ACM TASLP 27.8 (2019, arXiv 1809.07454) by Yi Luo and Nima Mesgarani: fully convolutional time-domain separation that surpassed ideal time-frequency masking on speech.
- · Author-reported 15.3 dB SI-SNRi on WSJ0-2mix; Asteroid reproduction reaches 16.2 dB. Template for DPRNN, SepFormer, and the time-domain family.
- · In a music catalog honestly: SDX'23 leaderboard lists it as a music-demixing baseline at 5.73 avg global SDR — baseline-grade, not competitive with music specialists.