Audio retrieval · Listening examples

Stembed

Stembed produces multiple slot embeddings for a mixture query containing multiple stems. Read the tags of each slot’s retrieved audio (e.g., “bass guitar”), then listen to the results.

View

In this evaluation, we select the top-m slots where m is the number of stems in the mixture. In a more realistic scenario, Stembed can rank the slots with a presence score and keep only slots whose presence score is above 50%. See the paper for more details.

Listen to each slot’s retrieved recordings. Annotated mode reveals the reference stems and their ranks. Model scores are available below each slot.

A green outline and checkmark identify an original recording. This demo annotation would be unavailable for an unknown mixture. Unmarked recordings may still be useful musical matches.

Three-stem mixtures

Search gallery: 2,068 stems. Three slots selected by presence.

Example 01bass guitar + distorted guitar + background vocals

Example 01
Listen to the mixture

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.226

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass synthesizer
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.226
  • Rank 2: Cosine distance 0.380
  • Rank 3: Cosine distance 0.398

Slot 3 · distorted guitar

Original rank 2 / 2,068
Cosine distance 0.233

  1. 1. distorted guitar
  2. 2. distorted guitarOriginal recording
  3. 3. distorted guitar
Model scores

Presence 99.5%

  • Rank 1: Cosine distance 0.154
  • Rank 2: Cosine distance 0.233
  • Rank 3: Cosine distance 0.240

Slot 1 · background vocals

Original rank 1 / 2,068
Cosine distance 0.189

  1. 1. background vocalsOriginal recording
  2. 2. background vocals
  3. 3. background vocals
Model scores

Presence 98.7%

  • Rank 1: Cosine distance 0.189
  • Rank 2: Cosine distance 0.263
  • Rank 3: Cosine distance 0.266
Listen to the isolated query stems
bass guitar
distorted guitar
background vocals

Example 02contrabass/double bass + male lead vocals + full acoustic drumkit

Example 02
Listen to the mixture

Slot 1 · male lead vocals

Original rank 23 / 2,068
Cosine distance 0.433

  1. 1. background vocals
  2. 2. male lead vocals
  3. 3. male lead vocals
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.274
  • Rank 2: Cosine distance 0.297
  • Rank 3: Cosine distance 0.309

Slot 4 · contrabass/double bass

Original rank 5 / 2,068
Cosine distance 0.367

  1. 1. contrabass/double bass
  2. 2. bass synthesizer
  3. 3. contrabass/double bass
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.319
  • Rank 2: Cosine distance 0.348
  • Rank 3: Cosine distance 0.356

Slot 3 · full acoustic drumkit

Original rank 5 / 2,068
Cosine distance 0.429

  1. 1. overheads
  2. 2. full acoustic drumkit
  3. 3. drum machine
Model scores

Presence 98.7%

  • Rank 1: Cosine distance 0.338
  • Rank 2: Cosine distance 0.388
  • Rank 3: Cosine distance 0.403
Listen to the isolated query stems
contrabass/double bass
male lead vocals
full acoustic drumkit

Example 03bass guitar + female lead vocals + snare drum

Example 03
Listen to the mixture

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.289

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.289
  • Rank 2: Cosine distance 0.303
  • Rank 3: Cosine distance 0.436

Slot 3 · snare drum

Original rank 1 / 2,068
Cosine distance 0.196

  1. 1. snare drumOriginal recording
  2. 2. snare drum
  3. 3. snare drum
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.196
  • Rank 2: Cosine distance 0.342
  • Rank 3: Cosine distance 0.342

Slot 1 · female lead vocals

Original rank 3 / 2,068
Cosine distance 0.404

  1. 1. female lead vocals
  2. 2. female lead vocals
  3. 3. female lead vocalsOriginal recording
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.329
  • Rank 2: Cosine distance 0.354
  • Rank 3: Cosine distance 0.404
Listen to the isolated query stems
bass guitar
female lead vocals
snare drum

Example 04bass guitar + clean guitar + hi hat

Example 04
Listen to the mixture

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.158

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.158
  • Rank 2: Cosine distance 0.232
  • Rank 3: Cosine distance 0.239

Slot 1 · clean guitar

Original rank 1 / 2,068
Cosine distance 0.181

  1. 1. clean guitarOriginal recording
  2. 2. clean guitar
  3. 3. clean guitar
Model scores

Presence 99.5%

  • Rank 1: Cosine distance 0.181
  • Rank 2: Cosine distance 0.249
  • Rank 3: Cosine distance 0.265

Slot 3 · hi hat

Original rank 1 / 2,068
Cosine distance 0.180

  1. 1. hi hatOriginal recording
  2. 2. a-tonal percussion
  3. 3. hi hat
Model scores

Presence 92.4%

  • Rank 1: Cosine distance 0.180
  • Rank 2: Cosine distance 0.354
  • Rank 3: Cosine distance 0.374
Listen to the isolated query stems
bass guitar
clean guitar
hi hat

Example 05contrabass/double bass + male lead vocals + electric piano

Example 05
Listen to the mixture

Slot 4 · contrabass/double bass

Original rank 1 / 2,068
Cosine distance 0.331

  1. 1. contrabass/double bassOriginal recording
  2. 2. bass guitar
  3. 3. contrabass/double bass
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.331
  • Rank 2: Cosine distance 0.348
  • Rank 3: Cosine distance 0.371

Slot 1 · male lead vocals

Original rank 4 / 2,068
Cosine distance 0.346

  1. 1. male lead vocals
  2. 2. background vocals
  3. 3. male lead vocals
Model scores

Presence 99.6%

  • Rank 1: Cosine distance 0.286
  • Rank 2: Cosine distance 0.307
  • Rank 3: Cosine distance 0.319

Slot 3 · electric piano

Original rank 4 / 2,068
Cosine distance 0.333

  1. 1. grand piano
  2. 2. grand piano
  3. 3. grand piano
Model scores

Presence 97.6%

  • Rank 1: Cosine distance 0.171
  • Rank 2: Cosine distance 0.298
  • Rank 3: Cosine distance 0.322
Listen to the isolated query stems
contrabass/double bass
male lead vocals
electric piano

Example 06bass guitar + clean guitar + background vocals

Example 06
Listen to the mixture

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.110

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.110
  • Rank 2: Cosine distance 0.351
  • Rank 3: Cosine distance 0.445

Slot 3 · clean guitar

Original rank 4 / 2,068
Cosine distance 0.422

  1. 1. clean guitar
  2. 2. clean guitar
  3. 3. clean guitar
Model scores

Presence 99.1%

  • Rank 1: Cosine distance 0.377
  • Rank 2: Cosine distance 0.396
  • Rank 3: Cosine distance 0.420

Slot 1 · background vocals

Original rank 6 / 2,068
Cosine distance 0.343

  1. 1. background vocals
  2. 2. male lead vocals
  3. 3. background vocals
Model scores

Presence 94.3%

  • Rank 1: Cosine distance 0.298
  • Rank 2: Cosine distance 0.307
  • Rank 3: Cosine distance 0.318
Listen to the isolated query stems
bass guitar
clean guitar
background vocals

Example 07background vocals + drum machine + electric piano

Example 07
Listen to the mixture

Slot 1 · background vocals

Original rank 14 / 2,068
Cosine distance 0.428

  1. 1. male lead vocals
  2. 2. background vocals
  3. 3. background vocals
Model scores

Presence 97.8%

  • Rank 1: Cosine distance 0.294
  • Rank 2: Cosine distance 0.356
  • Rank 3: Cosine distance 0.360

Slot 4 · electric piano

Original rank 2 / 2,068
Cosine distance 0.321

  1. 1. clean guitar
  2. 2. electric pianoOriginal recording
  3. 3. electric piano
Model scores

Presence 97.3%

  • Rank 1: Cosine distance 0.312
  • Rank 2: Cosine distance 0.321
  • Rank 3: Cosine distance 0.327

Slot 3 · drum machine

Original rank 1 / 2,068
Cosine distance 0.090

  1. 1. drum machineOriginal recording
  2. 2. drum machine
  3. 3. overheads
Model scores

Presence 92.9%

  • Rank 1: Cosine distance 0.090
  • Rank 2: Cosine distance 0.269
  • Rank 3: Cosine distance 0.339
Listen to the isolated query stems
background vocals
drum machine
electric piano

Example 08bass guitar + distorted guitar + female lead vocals

Example 08
Listen to the mixture

Slot 1 · female lead vocals

Original rank 1 / 2,068
Cosine distance 0.233

  1. 1. female lead vocalsOriginal recording
  2. 2. female lead vocals
  3. 3. background vocals
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.233
  • Rank 2: Cosine distance 0.314
  • Rank 3: Cosine distance 0.354

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.156

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 99.6%

  • Rank 1: Cosine distance 0.156
  • Rank 2: Cosine distance 0.277
  • Rank 3: Cosine distance 0.328

Slot 3 · distorted guitar

Original rank 41 / 2,068
Cosine distance 0.559

  1. 1. distorted guitar
  2. 2. distorted guitar
  3. 3. clean guitar
Model scores

Presence 95.2%

  • Rank 1: Cosine distance 0.274
  • Rank 2: Cosine distance 0.314
  • Rank 3: Cosine distance 0.360
Listen to the isolated query stems
bass guitar
distorted guitar
female lead vocals

Example 09bass synthesizer + male lead vocals + grand piano

Example 09
Listen to the mixture

Slot 4 · bass synthesizer

Original rank 1 / 2,068
Cosine distance 0.238

  1. 1. bass synthesizerOriginal recording
  2. 2. bass synthesizer
  3. 3. bass guitar
Model scores

Presence 99.6%

  • Rank 1: Cosine distance 0.238
  • Rank 2: Cosine distance 0.383
  • Rank 3: Cosine distance 0.394

Slot 1 · male lead vocals

Original rank 1 / 2,068
Cosine distance 0.152

  1. 1. male lead vocalsOriginal recording
  2. 2. male lead vocals
  3. 3. male lead vocals
Model scores

Presence 99.6%

  • Rank 1: Cosine distance 0.152
  • Rank 2: Cosine distance 0.204
  • Rank 3: Cosine distance 0.282

Slot 3 · grand piano

Original rank 1 / 2,068
Cosine distance 0.237

  1. 1. grand pianoOriginal recording
  2. 2. electric piano
  3. 3. grand piano
Model scores

Presence 98.4%

  • Rank 1: Cosine distance 0.237
  • Rank 2: Cosine distance 0.362
  • Rank 3: Cosine distance 0.362
Listen to the isolated query stems
bass synthesizer
male lead vocals
grand piano

Example 10bass guitar + acoustic guitar + female lead vocals

Example 10
Listen to the mixture

Slot 1 · female lead vocals

Original rank 27 / 2,068
Cosine distance 0.580

  1. 1. female lead vocals
  2. 2. female lead vocals
  3. 3. female lead vocals
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.242
  • Rank 2: Cosine distance 0.266
  • Rank 3: Cosine distance 0.353

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.086

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.086
  • Rank 2: Cosine distance 0.145
  • Rank 3: Cosine distance 0.228

Slot 3 · acoustic guitar

Original rank 1 / 2,068
Cosine distance 0.158

  1. 1. acoustic guitarOriginal recording
  2. 2. acoustic guitar
  3. 3. acoustic guitar
Model scores

Presence 99.6%

  • Rank 1: Cosine distance 0.158
  • Rank 2: Cosine distance 0.261
  • Rank 3: Cosine distance 0.273
Listen to the isolated query stems
bass guitar
acoustic guitar
female lead vocals

Example 11bass guitar + distorted guitar + background vocals

Example 11
Listen to the mixture

Slot 4 · bass guitar

Original rank 2 / 2,068
Cosine distance 0.190

  1. 1. bass guitar
  2. 2. bass guitarOriginal recording
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.190
  • Rank 2: Cosine distance 0.190
  • Rank 3: Cosine distance 0.209

Slot 3 · distorted guitar

Original rank 1 / 2,068
Cosine distance 0.190

  1. 1. distorted guitarOriginal recording
  2. 2. distorted guitar
  3. 3. distorted guitar
Model scores

Presence 99.6%

  • Rank 1: Cosine distance 0.190
  • Rank 2: Cosine distance 0.275
  • Rank 3: Cosine distance 0.277

Slot 1 · background vocals

Original rank 11 / 2,068
Cosine distance 0.333

  1. 1. background vocals
  2. 2. background vocals
  3. 3. background vocals
Model scores

Presence 93.8%

  • Rank 1: Cosine distance 0.178
  • Rank 2: Cosine distance 0.208
  • Rank 3: Cosine distance 0.256
Listen to the isolated query stems
bass guitar
distorted guitar
background vocals

Example 12acoustic guitar + male lead vocals + overheads

Example 12
Listen to the mixture

Slot 1 · male lead vocals

Original rank 2 / 2,068
Cosine distance 0.297

  1. 1. male lead vocals
  2. 2. male lead vocalsOriginal recording
  3. 3. male lead vocals
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.267
  • Rank 2: Cosine distance 0.297
  • Rank 3: Cosine distance 0.407

Slot 4 · acoustic guitar

Original rank 1 / 2,068
Cosine distance 0.140

  1. 1. acoustic guitarOriginal recording
  2. 2. acoustic guitar
  3. 3. acoustic guitar
Model scores

Presence 99.5%

  • Rank 1: Cosine distance 0.140
  • Rank 2: Cosine distance 0.190
  • Rank 3: Cosine distance 0.293

Slot 3 · overheads

Original rank 1 / 2,068
Cosine distance 0.079

  1. 1. overheadsOriginal recording
  2. 2. full acoustic drumkit
  3. 3. overheads
Model scores

Presence 97.3%

  • Rank 1: Cosine distance 0.079
  • Rank 2: Cosine distance 0.283
  • Rank 3: Cosine distance 0.308
Listen to the isolated query stems
acoustic guitar
male lead vocals
overheads

Example 13bass guitar + male lead vocals + kick drum

Example 13
Listen to the mixture

Slot 1 · male lead vocals

Original rank 1 / 2,068
Cosine distance 0.153

  1. 1. male lead vocalsOriginal recording
  2. 2. male lead vocals
  3. 3. male lead vocals
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.153
  • Rank 2: Cosine distance 0.170
  • Rank 3: Cosine distance 0.175

Slot 4 · bass guitar

Original rank 14 / 2,068
Cosine distance 0.476

  1. 1. bass synthesizer
  2. 2. bass synthesizer
  3. 3. bass synthesizer
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.326
  • Rank 2: Cosine distance 0.341
  • Rank 3: Cosine distance 0.371

Slot 2 · kick drum

Original rank 2 / 2,068
Cosine distance 0.162

  1. 1. kick drum
  2. 2. kick drumOriginal recording
  3. 3. kick drum
Model scores

Presence 96.8%

  • Rank 1: Cosine distance 0.128
  • Rank 2: Cosine distance 0.162
  • Rank 3: Cosine distance 0.173
Listen to the isolated query stems
bass guitar
male lead vocals
kick drum

Example 14bass synthesizer + clean guitar + male lead vocals

Example 14
Listen to the mixture

Slot 4 · bass synthesizer

Original rank 1 / 2,068
Cosine distance 0.185

  1. 1. bass synthesizerOriginal recording
  2. 2. bass synthesizer
  3. 3. bass synthesizer
Model scores

Presence 99.7%

  • Rank 1: Cosine distance 0.185
  • Rank 2: Cosine distance 0.236
  • Rank 3: Cosine distance 0.295

Slot 1 · male lead vocals

Original rank 15 / 2,068
Cosine distance 0.459

  1. 1. female lead vocals
  2. 2. background vocals
  3. 3. male lead vocals
Model scores

Presence 99.5%

  • Rank 1: Cosine distance 0.312
  • Rank 2: Cosine distance 0.316
  • Rank 3: Cosine distance 0.335

Slot 3 · clean guitar

Original rank 1 / 2,068
Cosine distance 0.291

  1. 1. clean guitarOriginal recording
  2. 2. clean guitar
  3. 3. clean guitar
Model scores

Presence 77.0%

  • Rank 1: Cosine distance 0.291
  • Rank 2: Cosine distance 0.328
  • Rank 3: Cosine distance 0.435
Listen to the isolated query stems
bass synthesizer
clean guitar
male lead vocals

Example 15contrabass/double bass + male lead vocals + full acoustic drumkit

Example 15
Listen to the mixture

Slot 4 · contrabass/double bass

Original rank 1 / 2,068
Cosine distance 0.273

  1. 1. contrabass/double bassOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 99.8%

  • Rank 1: Cosine distance 0.273
  • Rank 2: Cosine distance 0.437
  • Rank 3: Cosine distance 0.454

Slot 1 · male lead vocals

Original rank 1 / 2,068
Cosine distance 0.159

  1. 1. male lead vocalsOriginal recording
  2. 2. male lead vocals
  3. 3. male lead vocals
Model scores

Presence 99.4%

  • Rank 1: Cosine distance 0.159
  • Rank 2: Cosine distance 0.259
  • Rank 3: Cosine distance 0.302

Slot 2 · full acoustic drumkit

Original rank 46 / 2,068
Cosine distance 0.600

  1. 1. full acoustic drumkit
  2. 2. full acoustic drumkit
  3. 3. kick drum
Model scores

Presence 99.3%

  • Rank 1: Cosine distance 0.271
  • Rank 2: Cosine distance 0.334
  • Rank 3: Cosine distance 0.343
Listen to the isolated query stems
contrabass/double bass
male lead vocals
full acoustic drumkit

Example 16bass guitar + female lead vocals + snare drum

Example 16
Listen to the mixture

Slot 3 · snare drum

Original rank 1 / 2,068
Cosine distance 0.385

  1. 1. snare drumOriginal recording
  2. 2. snare drum
  3. 3. snare drum
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.385
  • Rank 2: Cosine distance 0.395
  • Rank 3: Cosine distance 0.404

Slot 4 · bass guitar

Original rank 10 / 2,068
Cosine distance 0.575

  1. 1. bass guitar
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.268
  • Rank 2: Cosine distance 0.335
  • Rank 3: Cosine distance 0.485

Slot 1 · female lead vocals

Original rank 1 / 2,068
Cosine distance 0.154

  1. 1. female lead vocalsOriginal recording
  2. 2. female lead vocals
  3. 3. female lead vocals
Model scores

Presence 99.8%

  • Rank 1: Cosine distance 0.154
  • Rank 2: Cosine distance 0.336
  • Rank 3: Cosine distance 0.367
Listen to the isolated query stems
bass guitar
female lead vocals
snare drum

Example 17bass guitar + female lead vocals + snare drum

Example 17
Listen to the mixture

Slot 3 · snare drum

Original rank 1 / 2,068
Cosine distance 0.184

  1. 1. snare drumOriginal recording
  2. 2. snare drum
  3. 3. snare drum
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.184
  • Rank 2: Cosine distance 0.212
  • Rank 3: Cosine distance 0.227

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.178

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.178
  • Rank 2: Cosine distance 0.234
  • Rank 3: Cosine distance 0.298

Slot 1 · female lead vocals

Original rank 1 / 2,068
Cosine distance 0.266

  1. 1. female lead vocalsOriginal recording
  2. 2. male lead vocals
  3. 3. male lead vocals
Model scores

Presence 99.7%

  • Rank 1: Cosine distance 0.266
  • Rank 2: Cosine distance 0.368
  • Rank 3: Cosine distance 0.374
Listen to the isolated query stems
bass guitar
female lead vocals
snare drum

Example 18clean guitar + female lead vocals + snare drum

Example 18
Listen to the mixture

Slot 1 · female lead vocals

Original rank 14 / 2,068
Cosine distance 0.493

  1. 1. male lead vocals
  2. 2. female lead vocals
  3. 3. female lead vocals
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.258
  • Rank 2: Cosine distance 0.309
  • Rank 3: Cosine distance 0.340

Slot 3 · snare drum

Original rank 1 / 2,068
Cosine distance 0.147

  1. 1. snare drumOriginal recording
  2. 2. snare drum
  3. 3. snare drum
Model scores

Presence 99.9%

  • Rank 1: Cosine distance 0.147
  • Rank 2: Cosine distance 0.208
  • Rank 3: Cosine distance 0.285

Slot 4 · clean guitar

Original rank 2 / 2,068
Cosine distance 0.259

  1. 1. distorted guitar
  2. 2. clean guitarOriginal recording
  3. 3. clean guitar
Model scores

Presence 98.9%

  • Rank 1: Cosine distance 0.203
  • Rank 2: Cosine distance 0.259
  • Rank 3: Cosine distance 0.317
Listen to the isolated query stems
clean guitar
female lead vocals
snare drum

Example 19bass guitar + bass synthesizer + kick drum

Example 19
Listen to the mixture

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.175

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.175
  • Rank 2: Cosine distance 0.254
  • Rank 3: Cosine distance 0.283

Slot 2 · kick drum

Original rank 1 / 2,068
Cosine distance 0.165

  1. 1. kick drumOriginal recording
  2. 2. kick drum
  3. 3. kick drum
Model scores

Presence 99.2%

  • Rank 1: Cosine distance 0.165
  • Rank 2: Cosine distance 0.239
  • Rank 3: Cosine distance 0.245

Slot 3 · bass synthesizer

Original rank 33 / 2,068
Cosine distance 0.609

  1. 1. acoustic guitar
  2. 2. acoustic guitar
  3. 3. acoustic guitar
Model scores

Presence 0.7%

  • Rank 1: Cosine distance 0.409
  • Rank 2: Cosine distance 0.447
  • Rank 3: Cosine distance 0.462
Listen to the isolated query stems
bass guitar
bass synthesizer
kick drum

Example 20bass guitar + distorted guitar + male lead vocals

Example 20
Listen to the mixture

Slot 4 · bass guitar

Original rank 1 / 2,068
Cosine distance 0.123

  1. 1. bass guitarOriginal recording
  2. 2. bass guitar
  3. 3. bass guitar
Model scores

Presence 100.0%

  • Rank 1: Cosine distance 0.123
  • Rank 2: Cosine distance 0.363
  • Rank 3: Cosine distance 0.422

Slot 1 · male lead vocals

Original rank 4 / 2,068
Cosine distance 0.312

  1. 1. background vocals
  2. 2. male lead vocals
  3. 3. male lead vocals
Model scores

Presence 95.0%

  • Rank 1: Cosine distance 0.224
  • Rank 2: Cosine distance 0.273
  • Rank 3: Cosine distance 0.306

Slot 3 · distorted guitar

Original rank 1 / 2,068
Cosine distance 0.161

  1. 1. distorted guitarOriginal recording
  2. 2. distorted guitar
  3. 3. distorted guitar
Model scores

Presence 93.5%

  • Rank 1: Cosine distance 0.161
  • Rank 2: Cosine distance 0.309
  • Rank 3: Cosine distance 0.327
Listen to the isolated query stems
bass guitar
distorted guitar
male lead vocals

What you are hearing

Each mixture produces four slots, each with a presence logit and a 256-dimensional embedding. At inference, the logits can be converted to probabilities with the sigmoid function, keeping embeddings whose probability exceeds 50%. Alternatively, the user can specify an expected stem count (m) and use the embeddings with the top m logits. This demo uses the latter: the top three slots from mixtures that always contain three stems.

We show annotations that would not ordinarily be available for unknown mixtures. We encode the ground-truth “solo” stems at the mixture’s query window. After selecting the top three slots by presence logit, we match their normalized embeddings one-to-one to these normalized solo embeddings using cosine distance. This assigns each selected slot a unique target for evaluation without changing its retrieval ranking. For example, one slot may be assigned to guitar and another to snare; each is then evaluated on retrieving its assigned stem from the gallery, usually at a different time window.

Green outlines indicate exact matches to one of the mixture’s constituent stems. A slot’s annotated recall counts only its assigned target. Gallery excerpts use a separate active window where possible.

Examples cycle through songs in the fixed test set, showing one mixture per song before repeating songs, without cherry-picking. Audio reproduces the evaluated waveform at its original level.

Cosine distance is 1 minus cosine similarity; lower is closer. Model scores show presence probabilities and candidate distances.

Three-stem model

MuQ layer 4, four decoder blocks, four learned queries, and 256-dimensional embeddings. Training uses 166 MoisesDB songs from an artist-disjoint train/validation/test split.

This page uses training seed 0 at validation-selected step 1,250, with normalized embeddings during training and matching. Checkpoint selection uses GT-assigned validation mean reciprocal rank. The paper reports averages across three training seeds using GT-assigned recall. No isolated references are needed for ordinary retrieval.

Hugging FacePrivate · access required

The 3-stem model on Hugging Face is the same model used for these listening examples. Installation and usage.