Mask-Based Speech Enhancement for Spatial Audio: A Comparison of Ambisonics, Beamforming, and Microphone Channels

1. INTRODUCTION

• Spatial-audio applications require interference suppression alongside preservation of cues important for localization, scene awareness, listener comfort, and spatial release from masking.Masking can distort inter-channel spatial information and degrade localization and spatial perception.Source: §1 · p. 1
• Prior comparisons of beamformer- and Ambisonics-domain masking did not systematically compare direct microphone-channel masking across speech and spatial-fidelity outcomes.The cited prior work found that Ambisonics masking can better preserve spatial information for both the desired signal and residual disturbance.Source: §1 · p. 1
• The study compares time–frequency masking on microphone channels (TFM), beamformer outputs (TFB), and Ambisonics (TFA), using direct and reverberant target references.It evaluates speech quality, intelligibility, spatial cue preservation, and reverberation characteristics in simulated scenes with a spherical microphone array.Source: §1 · p. 1
• The contributions include showing that a reverberant desired reference is needed to preserve reverberation and that beamformer masking best preserves speech quality and intelligibility.Ambisonics masking best preserves residual-disturbance spatial qualities, while all three representations preserve target spatial cues.Source: §1 · p. 1

2.1. Signal Model

• The microphone-domain signal is a Q × 1 frequency-domain vector containing the target’s direct and reverberant components plus additive noise and other interferers.The microphones are positioned on a rigid sphere, with positions specified by spherical coordinates.Source: §2.1 · p. 2
• The Ambisonics representation uses a vector of spherical-harmonic coefficients, while the beamformer representation contains outputs directed toward different look directions.The beamformer signal is a B × 1 vector, with direct, reverberant, and noise components.Source: §2.1 · p. 2

2.2. Transformation Between Signal Representations

• The Ambisonics domain is the baseline representation; linear decoding matrices produce microphone and beamformer representations, and an inverse transform may introduce small error.The decoding matrices are invertible in the study’s setup, though they may not be invertible in general.Source: §2.2 · p. 2

2.3. Time-Frequency Masks

• After STFT transformation, diagonal ideal ratio masks estimate the desired components, with the desired signal defined as either the direct or reverberant target.The masks require knowledge of true signal and noise components, so they are not realizable in practice and serve as theoretical upper-bound references.Source: §2.3 · p. 2

2.4. Binaural Representation

• Binaural signals are formed by applying the HRTF after transforming enhanced signals to the frequency domain.Beamformer- and microphone-domain signals are first transformed back to the spherical-harmonics domain.Source: §2.4 · p. 2

3. OBJECTIVE MEASURES OF PERFORMANCE

• SI-SDR, averaged across binaural channels, measures signal reconstruction against either the direct or reverberant target reference.Source: §3 · p. 2
• Absolute errors in ITD and ILD between processed and reference binaural signals quantify preservation of spatial cues.Source: §3 · p. 2
• The clarity measure C50 assesses reverberation, while DBSTOI measures binaural intelligibility and PESQ assesses speech quality.C50 compares early energy arriving within 0–50 ms with late energy; DBSTOI uses direct target speech as its reference, and PESQ is averaged across ears.Source: §3 · p. 3

4.1. Setup

• The Monte-Carlo study uses 100 simulated room realizations varying speech, room acoustics, SNR, DRR, and source geometry.Each realization evaluates the resulting acoustic configuration across SNRs, DRRs, and desired-signal definitions.Source: §4.1 · p. 3, Table 1 · p. 3

4.2. Methodology

• The methodology compares direct and reverberant desired-signal definitions, computes ideal ratio masks from known components, and applies them to microphone, beamformer, or Ambisonics signals.Ambisonics is truncated to order N = 3; beamformer and microphone signals are generated from Ambisonics representations.Source: §4.2 · p. 3

4.3. Results

• The results assess reverberation preservation, speech quality and intelligibility, and spatial attributes of both desired and interfering signals.The analysis uses the reverberant target as the desired signal in the following comparisons.Source: §4.3 · p. 3, §4.3.1 · p. 4

4.3.1. Preservation of Late Reverberation

• Using the direct target as the desired signal suppresses late reverberation and raises C50; with a reverberant target, TFA and TFM keep ∆C50 below the JND, while TFB is least accurate.The ∆C50 comparison uses the reverberant target as reference; the JND is approximately 1 dB.Source: §4.3.1 · pp. 3–4, Table 2 · p. 4

4.3.2. Preservation of Speech Quality and Intelligibility of the Desired Signal

• With a reverberant target reference, TFB achieves the highest SI-SDR, PESQ, and DBSTOI scores, while TFM has the lowest scores.The measures are averaged over the Monte-Carlo dataset.Source: §4.3.2 · p. 4, Table 3 · p. 4

4.3.3. Preservation of the Desired Signal Spatial Attributes

• For the reverberant target, all methods keep target ∆ITD and ∆ILD well below perceptual thresholds.The cited thresholds are about 100 µs for ITD and 1 dB for ILD.Source: §4.3.3 · p. 4, Table 3 · p. 4

4.3.4. Preservation of the Residual Interfering Signal Spatial Attributes

• TFA has the lowest interference-cue differences, especially for ∆ITD, and better preserves residual-interference spatial attributes than the other methods.All methods’ interference ∆ITD and ∆ILD exceed the just noticeable difference threshold.Source: §4.3.4 · p. 4, Table 4 · p. 4

• Beamformer channels provide superior noise suppression and speech quality, while Ambisonics channels better preserve residual-interference spatial attributes and acoustic-scene C50.The trade-off is between enhancement and spatial fidelity.Source: §5 · p. 4
• Microphone-channel masking is least effective because raw channels lack spatial separation.Source: §5 · p. 4
• All three representations preserve the target speaker’s ITD/ILD spatial cues within perceptual limits.Source: §5 · p. 4
• Defining the reverberant target as the desired signal significantly improves C50 preservation.Source: §5 · p. 4
• Future work should test practical DNN-estimated masks, listening-test outcomes, first-order Ambisonics, and non-ideal array configurations.Source: §5 · p. 4