Location-based Training with Complementary Folded Linear Orderings for Multichannel Speech Separation

This demo page lets you hear how location-based training (LBT) with different output topologies assign speakers to output channels according to azimuth, and how complementary folded linear orderings (CLBT-FLOs) improve two-speaker multichannel separation relative to circular ordering (LBT-CO).

Headphones are recommended. Playing one clip pauses the others. Paper is available at: arXiv:2609.12629

Contents

Use the sections in order. Each one presents one claim:

  1. ESTOI evaluation of selection policies — objective confirmation that evaluation under ESTOI matches SI-SDRi on PET-3 and URA-4+.
  2. Location-based speaker ordering — showing that the network follows speaker location in separated outputs (Male-Female Separation).
  3. Complementary FLO selection — which FLO is preferred depends on speaker-array geometry; a selected FLO outperforms LBT-CO (Male-Female Separation).
  4. Advantage across arrays (anechoic) — selected LBT-FLO produces higher quality than LBT-CO on PET-3 and URA-4+ arrays (Male-Male Separation).
  5. Advantage across arrays (reverberant) — the same comparison in a highly reverberant room (Female-Female Separation).

1. ESTOI Evaluation of Selection Policies

As stated in the paper, selection-policy trends are similar when the metric is ESTOI instead of SI-SDRi. The plots below are averaged over 16 speaker pairs for the PET-3 and URA-4+ arrays. The relative performance and ranking of selection policies under ESTOI are consistent with those observed using SI-SDRi.

ESTOI results for PET-3 array

PET-3 array

ESTOI results for URA-4+ array

URA-4+ array

2. Location-based Speaker Ordering (Azimuth Swap, Reverberant)

LBT assigns output channels by location. In 2.1 a male is at 0° and a female at 140°. In 2.2 the same utterances are placed at swapped azimuths. Array: PET-3; room: Simulated reverberant room (RT60=0.60s).

What to hear. After the swap, Speaker 1 should become the female at 0° and Speaker 2 the male at 140°. Compare the starred LBT-FLO (θ = 0°) with LBT-CO for speech quality and intelligibility. Listen for residual crosstalk (the other talker leaking into a channel), identity swaps between output channels.

2.1 Male at 0°, female at 140°
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Mixture spectrogram Speaker 1 Ground truth speaker 1 Speaker 1 (PET-3) LBT-CO speaker 1 Speaker 1 (PET-3) LBT-FLO 0 speaker 1 Speaker 1 (PET-3) LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (PET-3) LBT-CO speaker 2 Speaker 2 (PET-3) LBT-FLO 0 speaker 2 Speaker 2 (PET-3) LBT-FLO 90 speaker 2
2.2 Same utterances, azimuths swapped: female at 0°, male at 140°
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Swapped mixture spectrogram Speaker 1 Swapped ground truth speaker 1 Speaker 1 (PET-3) Swapped LBT-CO speaker 1 Speaker 1 (PET-3) Swapped LBT-FLO 0 speaker 1 Speaker 1 (PET-3) Swapped LBT-FLO 90 speaker 1
Speaker 2 Swapped ground truth speaker 2 Speaker 2 (PET-3) Swapped LBT-CO speaker 2 Speaker 2 (PET-3) Swapped LBT-FLO 0 speaker 2 Speaker 2 (PET-3) Swapped LBT-FLO 90 speaker 2

3. Complementary LBT-FLO Selection in Anechoic Conditions

Speaker 1 is fixed at 40° (male). When Speaker 2 is at 140°, LBT-FLO (θ = 0°) is selected; when Speaker 2 is at 300°, LBT-FLO (θ = 90°) is selected. Array: PET-3; anechoic.

What to hear. In 3.1 the θ = 0° outputs should be closer to the ground truth than LBT-CO and than θ = 90°. In 3.2 the preference reverses: θ = 90° should be the cleaner pair. A single FLO is not enough for all geometries.

3.1 Speaker 1 male at 40°, Speaker 2 female at 140°
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Anechoic mixture 40 and 140 Speaker 1 Ground truth speaker 1 Speaker 1 (PET-3) LBT-CO speaker 1 Speaker 1 (PET-3) LBT-FLO 0 speaker 1 Speaker 1 (PET-3) LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (PET-3) LBT-CO speaker 2 Speaker 2 (PET-3) LBT-FLO 0 speaker 2 Speaker 2 (PET-3) LBT-FLO 90 speaker 2
3.2 Speaker 1 male at 40°, Speaker 2 female at 300°
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) LBT-FLO (θ = 90°) ★
Mixture Anechoic mixture 40 and 300 Speaker 1 Ground truth speaker 1 Speaker 1 (PET-3) LBT-CO speaker 1 Speaker 1 (PET-3) LBT-FLO 0 speaker 1 Speaker 1 (PET-3) LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (PET-3) LBT-CO speaker 2 Speaker 2 (PET-3) LBT-FLO 0 speaker 2 Speaker 2 (PET-3) LBT-FLO 90 speaker 2

4. Consistent Advantage of CLBT-FLOs Across Array Geometries (Anechoic)

The same anechoic scene is inferred on PET-3 and URA-4+: Speaker 1 male at 0°, Speaker 2 male at 140°. Selected LBT-FLO is θ = 0° on both arrays.

What to hear. Both talkers are male, so location (not pitch) must keep the channels distinct. On each array, the starred FLO should show less crosstalk than LBT-CO. The mixture is the same; only the array changes.

4.1 PET-3 — male at 0°, male at 140°
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Anechoic mixture Speaker 1 Ground truth speaker 1 Speaker 1 (PET-3) PET-3 LBT-CO speaker 1 Speaker 1 (PET-3) PET-3 LBT-FLO 0 speaker 1 Speaker 1 (PET-3) PET-3 LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (PET-3) PET-3 LBT-CO speaker 2 Speaker 2 (PET-3) PET-3 LBT-FLO 0 speaker 2 Speaker 2 (PET-3) PET-3 LBT-FLO 90 speaker 2
4.2 URA-4+ — same scene
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Anechoic mixture Speaker 1 Ground truth speaker 1 Speaker 1 (URA-4+) URA LBT-CO speaker 1 Speaker 1 (URA-4+) URA LBT-FLO 0 speaker 1 Speaker 1 (URA-4+) URA LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (URA-4+) URA LBT-CO speaker 2 Speaker 2 (URA-4+) URA LBT-FLO 0 speaker 2 Speaker 2 (URA-4+) URA LBT-FLO 90 speaker 2

5. Consistent Advantage of CLBT-FLOs Across Array Geometries (Reverberant)

The same comparison in a simulated reverberant room: Speaker 1 female at 0°, Speaker 2 female at 140°, on PET-3 and URA-4+. Selected LBT-FLO is again θ = 0°.

What to hear. Reverberation makes leakage easier to notice. The starred FLO should keep each female more isolated than LBT-CO on both arrays, with less of the other talker in the background.

5.1 PET-3 — female at 0°, female at 140°
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Reverberant mixture Speaker 1 Ground truth speaker 1 Speaker 1 (PET-3) PET-3 LBT-CO speaker 1 Speaker 1 (PET-3) PET-3 LBT-FLO 0 speaker 1 Speaker 1 (PET-3) PET-3 LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (PET-3) PET-3 LBT-CO speaker 2 Speaker 2 (PET-3) PET-3 LBT-FLO 0 speaker 2 Speaker 2 (PET-3) PET-3 LBT-FLO 90 speaker 2
5.2 URA-4+ — same scene
Mixture Ground Truth LBT-CO (Baseline) LBT-FLO (θ = 0°) ★ LBT-FLO (θ = 90°)
Mixture Reverberant mixture Speaker 1 Ground truth speaker 1 Speaker 1 (URA-4+) URA LBT-CO speaker 1 Speaker 1 (URA-4+) URA LBT-FLO 0 speaker 1 Speaker 1 (URA-4+) URA LBT-FLO 90 speaker 1
Speaker 2 Ground truth speaker 2 Speaker 2 (URA-4+) URA LBT-CO speaker 2 Speaker 2 (URA-4+) URA LBT-FLO 0 speaker 2 Speaker 2 (URA-4+) URA LBT-FLO 90 speaker 2