Monaural Continuous-Radius Regional Speech Extraction with Cross-Radius Consistency Learning

1. INTRODUCTION

• Speaker-identity-based extraction often requires prior enrollment or target-specific information, whereas source-to-microphone distance selects speech by location rather than identity.Source: §1 · p. 1
• Prior monaural distance methods use fixed thresholds or local windows, while multichannel regional systems require arrays [9, 10, 11, 12], leaving continuous-radius monaural extraction underexplored.Source: §1 · p. 1
• The method extracts all speakers within a queried microphone-centered region, directly modeling cumulative targets with continuous-radius queries in one model.Source: §1 · p. 1
• Cross-radius consistency learning uses nested target-speaker sets to encourage consistent predictions between adjacent radii with the same nonempty set.Source: §1 · p. 1
• The proposed method outperforms fixed-threshold and local-range baselines across regional extraction tasks and remains effective with more speakers and additive noise.Source: §1 · p. 1

2.1. Problem formulation

• For a query radius up to 4 m, the target is the sum of speakers inside that radius, or silence if none are present; inference uses the user-specified radius without estimating source distances.Source distances are used to construct training and evaluation supervision.Source: §2.1 · p. 2

2.2. Extraction network with continuous-radius conditioning

• The network injects a continuous radius representation through FiLM between adjacent TF blocks and decodes the final block output into an estimated waveform.The extraction network uses a TF-GridNet backbone and an STFT representation.Source: §2.2 · pp. 2–3, Figure 2 · p. 2

2.3. Cross-radius consistency learning

• Cross-radius consistency compares adjacent sampled radii only when their nonempty target-speaker sets are unchanged, excluding pairs that cross speaker boundaries.The loss minimizes normalized waveform discrepancy for valid pairs; valid-pair selection uses source distances during training.Source: §2.3 · p. 3

2.4. Loss function

• For each query radius, nonempty targets use normalized reconstruction loss, while empty targets receive a weighted penalty that encourages silence; the extraction loss averages across radii.The total training loss also includes cross-radius consistency loss with weight λcon = 0.1.Source: §2.4 · p. 3

3.1. Data

• Training and validation use two-speaker mixtures with simulated RIRs; measured dEchorate RIRs provide the test set for real acoustic conditions.Speech comes from four corpora, and noise is sampled from DNS with 0.5 probability at 0–15 dB SNR.Source: §3.1 · p. 3

3.2. Baselines

• DSS uses fixed-radius models at 1–4 m, while QD-TSE outputs from 1-m local intervals are summed to approximate cumulative queries.The interval-summation baseline requires multiple forward passes; all methods are evaluated at 1, 2, 3, and 4 m.Source: §3.2 · p. 3

3.3. Evaluation metrics

• SI-SDR evaluates full-target preservation, SI-SDRi partial-target extraction, attenuation empty-region suppression, and RCE consistency on a 0.1-m grid.RCE is reported as the median over utterances containing adjacent query pairs with the same nonempty target-speaker set.Source: §3.3 · p. 3

3.4. Implementation details

• The network has approximately 0.66 M trainable parameters, and reported results are averaged over three runs with different random seeds.Training uses Adam for at most 50 epochs, with batch size 1 and a ReduceLROnPlateau scheduler.Source: §3.4 · p. 3

4.1. Comparison with baselines

• The proposed method leads all three regional metrics, beating DSS by 7.22 dB SI-SDR, 4.01 dB SI-SDRi, and 5.89 dB attenuation.Against interval-summed QD-TSE, gains are 3.26 dB SI-SDR, 3.28 dB SI-SDRi, and 7.91 dB attenuation; each estimate uses one forward pass rather than multiple local-range queries.Source: Table 1 · p. 4, §4.1 · p. 4

4.2. Ablation study

• Without CRC, continuous encoding raises SI-SDR from 24.68 dB to 27.04 dB and lowers RCE from 1.04% to 0.85%; continuous encoding with CRC is best on every metric.CRC improves both encoding schemes; with continuous encoding it raises SI-SDR by 2.21 dB, SI-SDRi by 1.32 dB, and attenuation by 10.88 dB, while reducing RCE to 0.46%.Source: Table 2 · p. 4, §4.2 · p. 4

4.3. Continuous-radius analysis

• Dense 0.1-m queries show stable outputs when the target-speaker set is unchanged and sharp changes at source boundaries, with strong empty-region suppression and preserved full-target SI-SDR.The aggregated results also show positive partial-region SI-SDRi across densely sampled radii.Source: Figure 4 · p. 4, §4.3 · p. 4

4.4. Evaluation under challenging conditions

• With three speakers, SI-SDR falls by 3.08 dB and SI-SDRi by 1.21 dB; at 10 dB SNR, they fall by 0.20 dB and 0.17 dB, respectively.Under the combined three-speaker, 10-dB condition, SI-SDRi is 3.10 dB, attenuation is 52.67 dB, and RCE is 4.15%.Source: Table 3 · p. 4, §4.4 · p. 4

• On measured RIRs, the method outperforms task-matched fixed-threshold and local-range baselines and remains effective with more speakers and additive noise.Source: §5 · p. 4
• Continuous radius encoding outperforms one-hot radius codes, while CRC improves extraction performance and cross-radius consistency.Source: §5 · p. 4
• Dense-query outputs remain stable when the target-speaker set is unchanged and change sharply at source boundaries.Source: §5 · p. 4
• The method directly models cumulative regional targets and supports queries within the trained range using a single model.Source: §5 · p. 4