New Generation of Music Source Separation
Music Source Separation for musicians
New Generation of Music Source Separation
음악가를 위한 Music Source Separation
Today we release a new generation of Music Source Separation model. This model breaks audio that has already been mixed down into material for re-creation. Using the tracks separated by each instrument that makes up the song, creators can change some of the melodies themselves, remix the song anew, or swap in new instruments. This model will be the finest tool with which artists can analyze music.
The DUETA Music Source Separation model splits a song into 14 tracks. Kick and Snare, Sub Bass and Mid Bass, Lead Vocal and Background Vocals, Pad and FX are each separated into their own track.
A new way of creating
Remix. Change only the Kick or build a new Drum Loop, and swap the Sub Bass for a different sound to make the song again. The rest of the tracks stay untouched.
Sampling. Take out just one role of the song, like the chords of the Pad or the rhythm of the Pluck, and use it as material for new work.
Analysis and referencing. Listen role by role to check how a reference track stacked its Background Vocals and how it divided Sub Bass and Mid Bass.
Practice and education. Take out the Lead Vocal and record a cover, or leave only the Bass line and learn that part. In a lesson, you can play how the drums are put together one part at a time.
Performance. Prepare DJ sets and live arrangements with the separated tracks.
Track separation for musicians
Source separation is the technology of dividing audio whose mix is finished back into its individual elements. Until now, most source separation divided a song into four branches: Drums, Vocals, Bass and Other. But for someone who wants to change the Kick, a track with the whole drum kit bundled together cannot be material, and taking a sound effect you like and using it as material for your next creation was also very hard. So we made a model that can separate out tracks that are useful to artists.
Listen
"Future Bass" by Music Unlimited from Pixabay, 1:50 to 2:20 of the song.
- Mix
DUETA
DUETA Music Source Separation, the model in production on 2026-10-23, default job, as delivered
- Kick
- Snare
- Hi-hat
- Ride
- Crash
- Perc
- Sub Bass
- Mid Bass
- Main Lead
- Pad
- Pluck
- Lead Vocal
- Background Vocals
- FX & Other
Moises
Moises Pro, custom separation with every instrument group selected (16 tracks), WAV export, 2026-10-23, as delivered
- Kick
- Snare
- Toms
- Hi-hat
- Cymbals
- Other kit
- Bass
- Acoustic Guitars
- Electric Guitars
- Piano
- Keys
- Strings
- Winds
- Vocals
- Background Vocals
- Other
Lalal.ai
Lalal.ai Pro, one pass per stem, WAV, as delivered: Andromeda for vocals, drums, bass, piano and guitar (run 2026-10-26); Phoenix for synthesizer, which Andromeda does not offer (run 2026-10-27)
- Vocals
- Drums
- Bass
- Piano
- Guitar
- Synth
MVSep
MVSep Ensemble All-In v.2025.06.30, Premium API, 32-bit float WAV, run 2026-10-26, as delivered
- Kick
- Snare
- Toms
- Hi-hat
- Ride
- Crash
- Bass
- Lead Vocals
- Back Vocals
- Other
Benchmark
We put the same songs into DUETA Music Source Separation, Lalal.ai, MVSep, Moises and Suno, and measured every track each product returned. The models used are as follows.
- Lalal.ai: Andromeda, separation by track type (Pro) [9]
- MVSep: Ensemble All-In v.2025.06.30 (Premium) [10]
- Moises: Pro, Hi-Fi custom separation, 16 tracks [11]
- Suno: Get Stems, Auto Split [8]
Metrics
Reproduction SDR. Measures how close the sound of all the returned tracks added together is to the original song.
Audio Fidelity (PQ). Measures how good each individual track sounds to a quality model (Production Quality, PQ, of Audiobox Aesthetics).
Separation Leakage. The more the same sound is contained in several tracks, the higher the value. Each track should hold only its own part, so lower is better. Separation Leakage is a metric defined to approximate the interference ratio of BSS Eval.
Test songs
For this test we made songs with Suno. The top 10 songs have almost no Lead Vocal or none at all, and the bottom 10 songs all have a Lead Vocal.
All songs were generated in Suno with the prompts below.
| Title | Genre | Prompt |
|---|---|---|
| Neon Promise | K-pop dance | K-pop dance, mixed group vocals, layered chorus harmonies, heavy kick, synth bass, claps, vocal chops, 120 bpm |
| Glass Hearts | K-pop dance | K-pop dance, girl group vocals with a rap break, bright synth stabs, punchy kick, sub bass, claps and finger snaps, 124 bpm |
| Concrete Halo | Hip hop | hip hop, confident male rap vocal, booming 808 sub bass, crisp trap hi-hats, sparse piano, vocal ad-libs, 92 bpm |
| Saturday Engine | Funk | funk, powerful female vocal, slap bass, brass section, clavinet, tight drums with congas and percussion, 110 bpm |
| Electric Feelings | Modern pop | modern pop, female lead vocal, layered backing harmonies, punchy kick and snare, deep sub bass, warm synth pad, bright plucks, 118 bpm |
| Fragile Light | Piano ballad | piano ballad, emotional female vocal, grand piano, string section, light drums entering in the chorus, 72 bpm |
| Wires and Weather | Indie rock | indie rock band, male vocal, live drums with ride and crash cymbals, two electric guitars, bass guitar, energetic chorus, 128 bpm |
| Rising Tides | Progressive house | progressive house, female vocal, big build and drop, supersaw lead, sidechained pad, risers and white noise FX, 126 bpm |
| Velvet Hour | R&B | R&B, smooth male vocal, layered backing vocals, Rhodes electric piano, warm bass, soft percussion and finger snaps, 84 bpm |
| Swing with Me | Jazz | jazz trio with vocal, female jazz singer, upright bass, brushed drums with ride cymbal, acoustic piano, 120 bpm swing |
| Alive Tonight | EDM festival house | EDM festival house, soaring female vocal hook through the whole song, big room kick, sidechained supersaw lead, pluck arpeggio, 128 bpm, short song, verse, build and vocal drop |
| Claim My Piece | Hip hop | hip hop, confident female rapper through the whole song with a sung hook, booming 808, crisp hi-hats, sparse piano, 92 bpm, short song, verse and hook |
| Don't Leave Yet | Dance pop (duet) | dance pop duet, male and female lead vocals trading lines and singing the chorus together, four-on-the-floor kick, funky bass guitar, bright synths, 120 bpm, short song, verse and chorus |
| Let the Day Begin | Pop ballad | modern pop ballad, emotional female lead vocal from start to end, piano, strings, soft drums, sub bass, 72 bpm, short song, verse and chorus |
| Run Until We Touch the Sky | Pop punk | pop punk rock, energetic male lead vocal with gang backing vocals, distorted guitars, driving drums, bass guitar, 170 bpm, short song, verse and chorus |
| Stay Until the Morning | Future bass | future bass, breathy female lead vocal with vocal chops, wide detuned synth chords, punchy drums, deep sub bass, 150 bpm, short song, verse and chorus |
| Straight From The Start | K-pop boy group | K-pop boy group, confident male lead vocal with a rap verse and a sung chorus, hard-hitting trap drums, 808 sub bass, synth brass stabs, 95 bpm, short song, verse and chorus |
| Watch Me Take The Whole Floor | K-pop dance (female vocal) | K-pop dance pop, powerful female lead vocal singing through the whole song with layered harmonies, punchy kick and snare, deep sub bass, bright synth lead, pluck arpeggio, warm pad, 124 bpm, short song, verse and big chorus |
| We Own Today | Indie pop | indie pop, bright female lead vocal singing throughout, jangly electric guitar, synth bass, live drums, handclaps, 118 bpm, short song, verse and chorus |
| Where the Buses Turn | R&B | contemporary R&B, smooth male lead vocal with stacked harmonies, Rhodes keys, slow trap drums, sub bass, 80 bpm, short song, verse and chorus |
Results
DUETA returned 14 tracks per song. The product that returned more is Moises (16), and the products that returned fewer are Lalal.ai (5), MVSep (10) and Suno (8.3).
| System | Tracks ↑ | Reproduction SDR ↑ (dB) | PQ ↑ | PAM ↑ | Leakage, group ↓ (%) | Leakage, other tracks ↓ (%) |
|---|---|---|---|---|---|---|
| DUETA | 14.0 | 55.7 | 7.24 | 0.583 | 4.0 | 10.7 |
| Lalal.ai | 5.0 | 64.0 | 7.12 | 0.598 | 9.1 | 11.2 |
| MVSep | 10.0 | 28.0 | 7.15 | 0.555 | 11.2 | 14.4 |
| Moises | 16.0 | 24.7 | 7.10 | 0.566 | 4.3 | 11.8 |
| Suno | 8.3 | 7.1 | 7.24 | 0.718 | 19.6 | 19.0 |
Table 1. The per-song average over all 20 songs. Track count, Reproduction SDR, PQ and PAM lead when higher, and Leakage leads when lower. Differences of less than 1 dB in Reproduction SDR, 0.5 pp in Leakage, 0.1 in PQ and 0.02 in PAM are regarded as a similar level.
| System | Reproduction SDR (dB), first 10 songs | Reproduction SDR (dB), 10 vocal songs | PQ, first 10 songs | PQ, 10 vocal songs | Leakage, group (%), first 10 songs | Leakage, group (%), 10 vocal songs | Leakage, other tracks (%), first 10 songs | Leakage, other tracks (%), 10 vocal songs |
|---|---|---|---|---|---|---|---|---|
| DUETA | 56.7 | 54.7 | 7.24 | 7.24 | 3.6 | 4.5 | 9.5 | 11.9 |
| Lalal.ai | 63.4 | 64.6 | 7.10 | 7.14 | 12.4 | 5.8 | 14.1 | 8.4 |
| MVSep | 27.9 | 28.1 | 7.18 | 7.11 | 16.6 | 5.8 | 14.7 | 14.2 |
| Moises | 24.2 | 25.2 | 7.11 | 7.09 | 4.5 | 4.0 | 10.9 | 12.8 |
| Suno | 8.7 | 5.5 | 7.23 | 7.25 | 19.7 | 19.5 | 18.6 | 19.5 |
Table 2. The first 10 songs and the 10 songs with vocals.
Separation Leakage
How it is calculated: for each of the four track groups (vocals, drums, bass and other sounds), we obtained the share of the group's energy that is explained by 'the original song minus that group', fitting it per STFT bin and per 8-second segment. It is a metric we defined to approximate the interference ratio of BSS Eval [6] without the original tracks, and the definition and code are provided on request.
In Separation Leakage comparing each track group with the rest of the original song, DUETA is 4.0%. The rows higher than DUETA are Lalal.ai (9.1%), MVSep (11.2%) and Suno (19.6%), and the similar row is Moises (4.3%). The first 10 songs are DUETA 3.6%, Lalal.ai 12.4%, MVSep 16.6%, Moises 4.5% and Suno 19.7%, and the 10 songs with vocals are DUETA 4.5%, Lalal.ai 5.8%, MVSep 5.8%, Moises 4.0% and Suno 19.5%.
Separation Leakage, against the same product's other tracks
Each returned track is fitted the same way, but instead of 'the original song minus the group', the same product's other tracks are used as the explaining signal. This is the form registered before scoring.
Reproduction SDR
It is 10·log10 of the energy of the original song divided by the energy of (the original song minus the sound of all the counted tracks added together), the SDR form of the Music Demixing Challenge [1] that takes the original song as the reference. For Lalal.ai, each separation was measured by adding the part and the remaining sound, then averaged over all separations.
Audio Fidelity
How it is calculated: each returned track was given the Production Quality (1 to 10) of Audiobox Aesthetics [2], and the track average was averaged again over the songs.
| System | CE (Content Enjoyment) | CU (Content Usefulness) | PC (Production Complexity) | PQ (Production Quality) |
|---|---|---|---|---|
| DUETA | 4.07 | 6.83 | 2.43 | 7.24 |
| Lalal.ai | 4.50 | 6.87 | 2.50 | 7.12 |
| MVSep | 4.07 | 6.80 | 2.38 | 7.15 |
| Moises | 4.10 | 6.73 | 2.56 | 7.10 |
| Suno | 5.04 | 7.01 | 3.34 | 7.24 |
Table 3. The four scores (1 to 10) of Audiobox Aesthetics, the song average over all 20 songs. The PQ of Figure 6 and Table 1 is Production Quality among them.
- CE (Content Enjoyment): How enjoyable it is to listen to. It looks at artistry and emotional delivery.
- CU (Content Usefulness): How useful it is as material for content production.
- PC (Production Complexity): How complexly the sound is built. It is low for one instrument and high when many sounds overlap. A higher score is not a better one.
- PQ (Production Quality): Technical sound quality. It looks at clarity, fidelity, noise and distortion, and frequency balance.
Since one separated track should hold only one sound, a low PC is natural, and a high one can be a sign that other sounds are mixed in. CE comes out low for every product, because one part heard alone is less enjoyable than the original song.
Try it
We hope the DUETA Music Source Separation model will be of direct help to artists, and we hope it becomes a tool with which you can gain inspiration as a creator and study music.
On the DUETA console you can easily upload a song and receive all of its tracks.
References
오늘 저희는 새 세대의 Music Source Separation 모델을 공개합니다. 이 모델은 이미 믹싱이 완료된 음원을 재창작의 재료로 분해합니다. 곡을 구성하는 악기별로 분리된 트랙을 이용하여 크리에이터가 직접 일부 선율을 변경하거나, 곡을 새롭게 리믹스하고, 새로운 악기로 바꿀 수 있습니다. 이 모델은 아티스트가 음악을 분석할 수 있는 가장 뛰어난 도구가 될 것입니다.
DUETA Music Source Separation 모델은 곡을 14개 트랙으로 나눕니다. Kick과 Snare, Sub Bass와 Mid Bass, Lead Vocal과 Background Vocals, Pad와 FX가 각자의 트랙으로 분리됩니다.
새로운 창작 방식
리믹스. Kick만 바꾸거나 Drum Loop를 새로 짜고, Sub Bass를 다른 소리로 바꿔 곡을 다시 만듭니다. 나머지 트랙은 손대지 않은 채로 남습니다.
샘플링. Pad의 화음이나 Pluck의 리듬처럼 곡의 한 역할만 떼어 내 새 작업의 재료로 씁니다.
분석 및 레퍼런싱. 레퍼런스 음원이 Background Vocals를 어떻게 쌓았는지, Sub Bass와 Mid Bass를 어떻게 나누었는지 역할별로 들으며 확인합니다.
연습과 교육. Lead Vocal을 빼고 커버를 녹음하거나, Bass 라인만 남겨 그 파트를 익힙니다. 레슨에서는 드럼이 어떻게 짜여 있는지 한 파트씩 들려줄 수 있습니다.
공연. 분리된 트랙으로 DJ 세트와 라이브 편곡을 준비합니다.
뮤지션을 위한 트랙 분리
음원 분리는 믹스가 끝난 음원을 각각의 요소로 다시 나누는 기술입니다. 지금까지의 음원 분리는 대부분 곡을 Drums, Vocals, Bass, Other의 네 갈래로 나누었습니다. 그러나 Kick을 바꾸려는 사람에게 드럼 전체가 묶인 트랙은 재료가 되지 못하고, 마음에 드는 효과음을 가져와 다음 창작의 재료로 쓰는 일도 매우 어려웠습니다. 그래서 저희는 아티스트에게 유용한 트랙들을 분리해낼 수 있는 모델을 만들었습니다.
들어 보기
"Future Bass" by Music Unlimited from Pixabay, 1:50 to 2:20 of the song.
- Mix
DUETA
DUETA Music Source Separation, the model in production on 2026-10-23, default job, as delivered
- Kick
- Snare
- Hi-hat
- Ride
- Crash
- Perc
- Sub Bass
- Mid Bass
- Main Lead
- Pad
- Pluck
- Lead Vocal
- Background Vocals
- FX & Other
Moises
Moises Pro, custom separation with every instrument group selected (16 tracks), WAV export, 2026-10-23, as delivered
- Kick
- Snare
- Toms
- Hi-hat
- Cymbals
- Other kit
- Bass
- Acoustic Guitars
- Electric Guitars
- Piano
- Keys
- Strings
- Winds
- Vocals
- Background Vocals
- Other
Lalal.ai
Lalal.ai Pro, one pass per stem, WAV, as delivered: Andromeda for vocals, drums, bass, piano and guitar (run 2026-10-26); Phoenix for synthesizer, which Andromeda does not offer (run 2026-10-27)
- Vocals
- Drums
- Bass
- Piano
- Guitar
- Synth
MVSep
MVSep Ensemble All-In v.2025.06.30, Premium API, 32-bit float WAV, run 2026-10-26, as delivered
- Kick
- Snare
- Toms
- Hi-hat
- Ride
- Crash
- Bass
- Lead Vocals
- Back Vocals
- Other
Benchmark
저희는 DUETA Music Source Separation, Lalal.ai, MVSep, Moises, Suno에 같은 곡을 넣고, 각 제품이 돌려준 모든 트랙을 측정했습니다. 사용한 모델은 다음과 같습니다.
- Lalal.ai: Andromeda, 트랙 종류별 분리(Pro) [9]
- MVSep: Ensemble All-In v.2025.06.30(Premium) [10]
- Moises: Pro, Hi-Fi 사용자 지정 분리, 트랙 16개 [11]
- Suno: Get Stems, Auto Split [8]
Metrics
Reproduction SDR. 돌려받은 모든 트랙을 더한 소리가 원곡과 얼마나 가까운지를 측정합니다.
Audio Fidelity(PQ). 트랙 하나하나가 품질 모델(Audiobox Aesthetics의 Production Quality, PQ)에 얼마나 좋게 들리는지를 측정합니다.
Separation Leakage. 같은 소리가 여러 트랙에 들어 있을수록 값이 높아집니다. 트랙마다 자기 파트만 담겨야 하므로 낮을수록 좋습니다. Separation Leakage는 BSS Eval의 간섭 비율을 근사하도록 정한 지표입니다.
테스트 곡
이번 테스트를 위해 Suno로 곡을 만들었습니다. 상단 10곡에는 Lead Vocal이 거의 없거나 아예 없고, 하단 10곡에는 모두 Lead Vocal이 있습니다.
모든 곡은 아래 프롬프트로 Suno에서 생성했습니다.
| 제목 | 장르 | 프롬프트 |
|---|---|---|
| Neon Promise | 케이팝 댄스 | K-pop dance, mixed group vocals, layered chorus harmonies, heavy kick, synth bass, claps, vocal chops, 120 bpm |
| Glass Hearts | 케이팝 댄스 | K-pop dance, girl group vocals with a rap break, bright synth stabs, punchy kick, sub bass, claps and finger snaps, 124 bpm |
| Concrete Halo | 힙합 | hip hop, confident male rap vocal, booming 808 sub bass, crisp trap hi-hats, sparse piano, vocal ad-libs, 92 bpm |
| Saturday Engine | 펑크 | funk, powerful female vocal, slap bass, brass section, clavinet, tight drums with congas and percussion, 110 bpm |
| Electric Feelings | 모던 팝 | modern pop, female lead vocal, layered backing harmonies, punchy kick and snare, deep sub bass, warm synth pad, bright plucks, 118 bpm |
| Fragile Light | 피아노 발라드 | piano ballad, emotional female vocal, grand piano, string section, light drums entering in the chorus, 72 bpm |
| Wires and Weather | 인디 록 | indie rock band, male vocal, live drums with ride and crash cymbals, two electric guitars, bass guitar, energetic chorus, 128 bpm |
| Rising Tides | 프로그레시브 하우스 | progressive house, female vocal, big build and drop, supersaw lead, sidechained pad, risers and white noise FX, 126 bpm |
| Velvet Hour | R&B | R&B, smooth male vocal, layered backing vocals, Rhodes electric piano, warm bass, soft percussion and finger snaps, 84 bpm |
| Swing with Me | 재즈 | jazz trio with vocal, female jazz singer, upright bass, brushed drums with ride cymbal, acoustic piano, 120 bpm swing |
| Alive Tonight | EDM 페스티벌 하우스 | EDM festival house, soaring female vocal hook through the whole song, big room kick, sidechained supersaw lead, pluck arpeggio, 128 bpm, short song, verse, build and vocal drop |
| Claim My Piece | 힙합 | hip hop, confident female rapper through the whole song with a sung hook, booming 808, crisp hi-hats, sparse piano, 92 bpm, short song, verse and hook |
| Don't Leave Yet | 댄스 팝(듀엣) | dance pop duet, male and female lead vocals trading lines and singing the chorus together, four-on-the-floor kick, funky bass guitar, bright synths, 120 bpm, short song, verse and chorus |
| Let the Day Begin | 팝 발라드 | modern pop ballad, emotional female lead vocal from start to end, piano, strings, soft drums, sub bass, 72 bpm, short song, verse and chorus |
| Run Until We Touch the Sky | 팝 펑크 | pop punk rock, energetic male lead vocal with gang backing vocals, distorted guitars, driving drums, bass guitar, 170 bpm, short song, verse and chorus |
| Stay Until the Morning | 퓨처 베이스 | future bass, breathy female lead vocal with vocal chops, wide detuned synth chords, punchy drums, deep sub bass, 150 bpm, short song, verse and chorus |
| Straight From The Start | 케이팝 보이 그룹 | K-pop boy group, confident male lead vocal with a rap verse and a sung chorus, hard-hitting trap drums, 808 sub bass, synth brass stabs, 95 bpm, short song, verse and chorus |
| Watch Me Take The Whole Floor | 케이팝 댄스(여성 보컬) | K-pop dance pop, powerful female lead vocal singing through the whole song with layered harmonies, punchy kick and snare, deep sub bass, bright synth lead, pluck arpeggio, warm pad, 124 bpm, short song, verse and big chorus |
| We Own Today | 인디 팝 | indie pop, bright female lead vocal singing throughout, jangly electric guitar, synth bass, live drums, handclaps, 118 bpm, short song, verse and chorus |
| Where the Buses Turn | R&B | contemporary R&B, smooth male lead vocal with stacked harmonies, Rhodes keys, slow trap drums, sub bass, 80 bpm, short song, verse and chorus |
결과
DUETA는 곡마다 14개의 트랙을 돌려주었습니다. 더 많이 돌려준 제품은 Moises(16)이고, 더 적게 돌려준 제품은 Lalal.ai(5), MVSep(10), Suno(8.3)입니다.
| 시스템 | 트랙 수 ↑ | Reproduction SDR ↑ (dB) | PQ ↑ | PAM ↑ | Leakage, 그룹 ↓ (%) | Leakage, 다른 트랙 ↓ (%) |
|---|---|---|---|---|---|---|
| DUETA | 14.0 | 55.7 | 7.24 | 0.583 | 4.0 | 10.7 |
| Lalal.ai | 5.0 | 64.0 | 7.12 | 0.598 | 9.1 | 11.2 |
| MVSep | 10.0 | 28.0 | 7.15 | 0.555 | 11.2 | 14.4 |
| Moises | 16.0 | 24.7 | 7.10 | 0.566 | 4.3 | 11.8 |
| Suno | 8.3 | 7.1 | 7.24 | 0.718 | 19.6 | 19.0 |
표 1. 20곡 전체의 곡 평균입니다. 트랙 수, Reproduction SDR, PQ, PAM은 높을수록, Leakage는 낮을수록 앞섭니다. Reproduction SDR 1 dB, Leakage 0.5%p, PQ 0.1, PAM 0.02 미만의 차이는 비슷한 수준으로 봅니다.
| 시스템 | Reproduction SDR (dB), 처음 10곡 | Reproduction SDR (dB), 보컬 10곡 | PQ, 처음 10곡 | PQ, 보컬 10곡 | Leakage, 그룹 (%), 처음 10곡 | Leakage, 그룹 (%), 보컬 10곡 | Leakage, 다른 트랙 (%), 처음 10곡 | Leakage, 다른 트랙 (%), 보컬 10곡 |
|---|---|---|---|---|---|---|---|---|
| DUETA | 56.7 | 54.7 | 7.24 | 7.24 | 3.6 | 4.5 | 9.5 | 11.9 |
| Lalal.ai | 63.4 | 64.6 | 7.10 | 7.14 | 12.4 | 5.8 | 14.1 | 8.4 |
| MVSep | 27.9 | 28.1 | 7.18 | 7.11 | 16.6 | 5.8 | 14.7 | 14.2 |
| Moises | 24.2 | 25.2 | 7.11 | 7.09 | 4.5 | 4.0 | 10.9 | 12.8 |
| Suno | 8.7 | 5.5 | 7.23 | 7.25 | 19.7 | 19.5 | 18.6 | 19.5 |
표 2. 처음 10곡과 보컬이 있는 10곡입니다.
Separation Leakage
계산 방법: 네 트랙 그룹(보컬, 드럼, 베이스, 그 밖의 소리)마다 그룹의 에너지 가운데 '원곡에서 그 그룹을 뺀 소리'로 설명되는 비율을 STFT 빈과 8초 구간마다 맞춰 구했습니다. 원본 트랙 없이 BSS Eval [6]의 간섭 비율을 근사하도록 저희가 정한 지표이며, 정의와 코드는 요청하시면 제공합니다.
트랙 그룹마다 원곡의 나머지 소리와 비교한 Separation Leakage에서 DUETA는 4.0%입니다. DUETA보다 높은 행은 Lalal.ai(9.1%), MVSep(11.2%), Suno(19.6%)이고, 비슷한 행은 Moises(4.3%)입니다. 처음 10곡은 DUETA 3.6%, Lalal.ai 12.4%, MVSep 16.6%, Moises 4.5%, Suno 19.7%이고, 보컬이 있는 10곡은 DUETA 4.5%, Lalal.ai 5.8%, MVSep 5.8%, Moises 4.0%, Suno 19.5%입니다.
Separation Leakage, 같은 제품의 다른 트랙 기준
돌려받은 트랙마다 같은 방식으로 맞추되, '원곡에서 그룹을 뺀 소리' 대신 같은 제품의 다른 트랙을 설명 신호로 씁니다. 채점 전에 등록해 둔 형태입니다.
Reproduction SDR
원곡의 에너지를 (원곡에서 센 트랙을 모두 더한 소리를 뺀 신호)의 에너지로 나눈 값의 10·log10이며, 원곡을 기준으로 삼은 Music Demixing Challenge [1]의 SDR 형태입니다. Lalal.ai는 분리마다 파트와 나머지 소리를 더해 잰 뒤 모든 분리에 걸쳐 평균했습니다.
Audio Fidelity
계산 방법: 돌려받은 트랙마다 Audiobox Aesthetics [2]의 Production Quality(1~10)를 매기고, 트랙 평균을 다시 곡에 걸쳐 평균했습니다.
| 시스템 | CE (Content Enjoyment) | CU (Content Usefulness) | PC (Production Complexity) | PQ (Production Quality) |
|---|---|---|---|---|
| DUETA | 4.07 | 6.83 | 2.43 | 7.24 |
| Lalal.ai | 4.50 | 6.87 | 2.50 | 7.12 |
| MVSep | 4.07 | 6.80 | 2.38 | 7.15 |
| Moises | 4.10 | 6.73 | 2.56 | 7.10 |
| Suno | 5.04 | 7.01 | 3.34 | 7.24 |
표 3. Audiobox Aesthetics의 네 점수(1~10), 20곡 전체의 곡 평균입니다. 그림 6과 표 1의 PQ는 이 중 Production Quality입니다.
- CE (Content Enjoyment): 들어서 즐거운 정도입니다. 예술성과 감정 전달을 봅니다.
- CU (Content Usefulness): 콘텐츠 제작 재료로 쓸모 있는 정도입니다.
- PC (Production Complexity): 소리가 얼마나 복잡하게 짜였는지입니다. 악기 하나면 낮고, 여러 소리가 겹치면 높습니다. 높을수록 좋은 점수가 아닙니다.
- PQ (Production Quality): 기술적 음질입니다. 깨끗함, 충실도, 잡음과 왜곡, 주파수 균형을 봅니다.
분리된 트랙 하나에는 소리 하나만 있어야 하므로 PC는 낮은 편이 자연스럽고, 높으면 다른 소리가 섞였다는 신호일 수 있습니다. CE는 한 파트만 따로 들으면 원곡보다 덜 즐거워서 모든 제품에서 낮게 나옵니다.
사용해 보기
저희는 Dueta Music Separation Model이 아티스트분들에게 직접적인 도움이 되길 바라며, 창작자로써의 영감을 얻고 음악을 공부할 수 있는 도구가 되기를 바랍니다.
DUETA 콘솔에서 손쉽게 곡을 업로드하여, 전체 트랙을 받아 보실 수 있습니다.