Skip to content
ModelListedExternal
Download

ConvTasNet Lyrics NonCausal

A NonCausal Lyrics/Accompaniment separation model for the CAD2-Task1 baseline system.

Sourcehuggingface.co/cadenzachallenge/ConvTasNet_Lyrics_NonCausal/tree/96a3d10364260d1de62b70b8e55f1e5d9e71fb9a
KindOtherVersionRevision 96a3d10

Model source

Kind
Other
File size
41.4–41.5 MB
Version
vRevision 96a3d10
Updated
Sep 18, 2026
Downloads
24
Tasks
Audio
Source
Hugging Face

Source description

Cadenza Challenge: CAD2-Task1

A NonCausal Lyrics/Accompaniment separation model for the CAD2-Task1 baseline system.

  • Architecture: ConvTasNet (Kaituo XU) with multichannel support (Alexandre Defossez).
  • Parameters:
    • B: 256
    • C: 2
    • H: 512
    • L: 20
    • N: 256
    • P: 3
    • R: 4
    • X: 10
    • audio_channels: 2
    • causal: false
    • mask_nonlinear: relu
    • norm_type: gLN
  • training:
    • sample_rate: 44100
    • samples_per_track: 64
    • segment: 5.0
    • aggregate: 2
    • batch_size: 4
    • early_stop: true
    • epochs: 200

Dataset

The model was trained on the training split of the MUSDB18-HQ dataset.

How to use

from tasnet import ConvTasNetStereo

model = ConvTasNetStereo.from_pretrained(
    "cadenzachallenge/ConvTasNet_LyricsSeparation_NonCausal"
).cpu()

Results

TrackVocals (SDR)Accompaniment (SDR)
Al James - Schoolboy Facination6.8419.074
AM Contra - Heart Peripheral6.94814.105
Angels In Amplifiers - I'm Alright7.35810.859
Arise - Run Run Run6.10516.806
Ben Carrigan - We'll Talk About It All Tonight2.85310.181
BKS - Bulldozer1.90913.944
BKS - Too Much8.61513.212
Bobby Nobody - Stitch Up7.94812.685
Buitraker - Revo X4.60914.61
Carlos Gonzalez - A Place For Us4.2358.888
Cristina Vane - So Easy8.75913.639
Detsky Sad - Walkie Talkie7.73210.844
Enda Reilly - Cur An Long Ag Seol9.60313.723
Forkupines - Semantics4.95511.561
Georgia Wonder - Siren4.1248.578
Girls Under Glass - We Feel Alright4.3812.272
Hollow Ground - Ill Fate7.04616.299
James Elder & Mark M Thompson - The English Actor4.6949.638
Juliet's Rescue - Heartbeats6.28114.409
Little Chicago's Finest - My Own6.3136.603
Louis Cressy Band - Good Time6.50112.016
Lyndsey Ollard - Catching Up9.1812.116
M.E.R.C. Music - Knockout6.6198.507
Moosmusic - Big Dummy Shake8.09714.578
Motor Tapes - Shore0.76910.137
Mu - Too Bright5.85313.135
Nerve 9 - Pray For The Rain6.42514.427
PR - Happy Daze051.092
PR - Oh No09.021
Punkdisco - Oral Hygiene5.72517.681
Raft Monk - Tiring2.3789.244
Sambasevam Shanmugam - Kaathaadi8.16410.588
Secretariat - Borderline5.52210.817
Secretariat - Over The Top7.85914.996
Side Effects Project - Sing With Me11.19712.63
Signe Jakobsen - What Have You Done To Me7.68511.013
Skelpolu - Resurrection07.603
Speak Softly - Broken Man3.99714.516
Speak Softly - Like Horses6.4629.426
The Doppler Shift - Atrophy0.71114.241
The Easton Ellises - Falcon 692.4017.889
The Easton Ellises (Baumi) - SDRNR1.4797.948
The Long Wait - Dark Horses6.5312.661
The Mountaineering Club - Mallory10.66515.311
The Sunshine Garcia Band - For I Am The Moon9.59113.297
Timboz - Pony4.02514.271
Tom McKenzie - Directions8.03116.129
Triviul feat. The Fiend - Widow7.0618.168
We Fell From The Sky - Not You3.86211.685
Zeno - Signs6.36411.552
Total (median over frames, median over tracks)6.33812.194

Sources

1 source
Hugging FacePrimaryVersion 96a3d10364260d1de62b70b8e55f1e5d9e71fb9a
huggingface.co/cadenzachallenge/ConvTasNet_Lyrics_NonCausal/tree/96a3d10364260d1de62b70b8e55f1e5d9e71fb9a

Verified Sep 18

Model artifacts

2 artifacts

model.safetensors

safetensors · 41.4 MB · SHA-256 38ab32d382f9…a9e1 · Hugging Face

Download

pytorch_model.bin

bin · 41.5 MB · SHA-256 4592b6df0dae…dc0e · Hugging Face

Download

Source excerpts

2 excerpts
ReadmeSep 18huggingface.co/cadenzachallenge/ConvTasNet_Lyrics_NonCausal

--- license: apache-2.0 language: - en tags: - hearing loss - challenge - signal processing - source separation - lyrics intelligibility - audio - audio-to-audio --- # Cadenza Challenge: CAD2-Task1 A NonCausal Lyrics/Accompaniment separation model for the CAD2-Task1 baseline system. * Architecture: ConvTasNet (Kaituo XU) with multichannel support (Alexandre Defossez). * Parameters: * B: 256 * C: 2 * H: 512 * L: 20 * N: 256 * P: 3 * R: 4 * X: 10 * audio_channels: 2 * causal: false * mask_nonlinear: relu * norm_type: gLN * training: * sample_rate: 44100 * samples_per_track: 64 * segment: 5.0 * aggregate: 2 * batch_size: 4 * early_stop: true * epochs: 200 ## Dataset The model was trained on the training split of the MUSDB18-HQ dataset. ## How to use ``` from tasnet import ConvTasNetStereo model = ConvTasNetStereo.from_pretrained( "cadenzachallenge/ConvTasNet_LyricsSeparation_NonCausal" ).cpu() ``` ## Results | Track | Vocals (SDR) | Accompaniment (SDR) | |:------|:------------:|:---------:| | Al James - Schoolboy Facination | 6.841 | 9.074 | | AM Contra - Heart Peripheral | 6.948 | 14.105 | | Angels In Amplifiers - I'm Alright | 7.358 | 10.859 | | Arise - Run Run Run | 6.105 | 16.806 | | Ben Carrigan - We'll Talk About It All Tonight | 2.853 | 10.181 | | BKS - Bulldozer | 1.909 | 13.944 | | BKS - Too Much | 8.615 | 13.212 | | Bobby Nobody - Stitch Up | 7.948 | 12.685 | | Buitraker - Revo X | 4.609 | 14.61 | | Carlos Gonzalez - A Place For Us | 4.235 | 8.888 | | Cristina Vane - So Easy | 8.759 | 13.639 | | Detsky Sad - Walkie Talkie | 7.732 | 10.844 | | Enda Reilly - Cur An Long Ag Seol | 9.603 | 13.723 | | Forkupines - Semantics | 4.955 | 11.561 | | Georgia Wonder - Siren | 4.124 | 8.578 | | Girls Under Glass - We Feel Alright | 4.38 | 12.272 | | Hollow Ground - Ill Fate | 7.046 | 16.299 | | James Elder & Mark M Thompson - The English Actor | 4.694 | 9.638 | | Juliet's Rescue - Heartbeats | 6.281 | 14.409 | | Little Chicago's Finest - My Own | 6.313 | 6.603 | | Louis Cressy Band - Good Time | 6.501 | 12.016 | | Lyndsey Ollard - Catching Up | 9.18 | 12.116 | | M.E.R.C. Music - Knockout | 6.619 | 8.507 | | Moosmusic - Big Dummy Shake | 8.097 | 14.578 | | Motor Tapes - Shore | 0.769 | 10.137 | | Mu - Too Bright | 5.853 | 13.135 | | Nerve 9 - Pray For The Rain | 6.425 | 14.427 | | PR - Happy Daze | 0 | 51.092 | | PR - Oh No | 0 | 9....

NoteSep 18huggingface.co/cadenzachallenge/ConvTasNet_Lyrics_NonCausal

Source context: 24 downloads · 0 likes · Pipeline audio-to-audio · Library transformers · Repo cadenzachallenge/ConvTasNet_Lyrics_NonCausal