Source-linked AI summary
Frequency-Weighted Soft-Constrained Spatially Selective Active Noise Control for Open-Fitting Hearables
Tong Xiao, Reinhild Roden, Matthias Blau, Simon Doclo
TL;DR
Soft-constrained SSANC for open-fitting hearables uses a scalar trade-off across frequencies, while frequency-dependent weighting may conflict with low-delay control. This paper introduces frequency-dependent speech-preservation weighting in time-domain SSANC; LTASS and oracle yield the strongest reported intelligibility improvements at lower distortion than SII and MIRS.
Problem
Soft-constrained SSANC balances noise reduction and desired-speech preservation with a scalar across frequencies, while frequency-dependent alternatives are typically block-based and may conflict with low-delay control.
Source: §1 · p. 1Method
The paper generalizes soft-constrained SSANC with frequency-dependent desired-speech preservation weights while maintaining low-delay time-domain active control.
Source: §1 · p. 1Results
Takeaways & Limitations
LTASS offers a perceptual trade-off comparable to oracle weighting without requiring the clean desired-speech signal.
Source: §1 · p. 1The signal-model analysis assumes an exact secondary-path estimate, making its modeled leakage estimate equal to the leakage signal.
Source: §2 · p. 2
Abstract
from arXiv · showhide
Recently, soft-constrained spatially selective active noise control (SSANC) for open-fitting hearables has been proposed to enable a flexible trade-off between noise reduction and desired-speech preservation using a scalar trade-off parameter across all frequencies. This paper generalizes soft-constrained SSANC by introducing frequency-dependent weighting of the desired-speech preservation error while maintaining low-delay time-domain active control. Frequency-dependent weightings based on the speech intelligibility index (SII), the modified intermediate reference system (MIRS), the long-term average speech spectrum (LTASS), and a signal-dependent oracle are compared with scalar weighting using measured acoustic impulse responses in a multi-speaker scenario. The results show that LTASS and oracle weightings provide the largest speech-intelligibility improvements and comparable speech-quality improvements at substantially lower distortion levels than those obtained with SII and MIRS weightings. While the oracle weighting provides the best overall perceptual trade-off, LTASS weighting provides a comparable perceptual trade-off without requiring the clean desired-speech signal.
1. INTRODUCTION
Open-fitting hearables can preserve a desired speaker while attenuating other sounds, but acoustic leakage reaches the eardrum. This work extends soft-constrained SSANC with frequency-dependent speech-preservation weighting while retaining low-delay time-domain control.
1. INTRODUCTION
- Open-fitting hearables improve comfort and reduce occlusion, but allow desired and undesired sounds to leak to the eardrum.Source: §1 · p. 1
- SSANC combines spatial filtering with active control to attenuate noise while preserving desired speech at the inner microphone.Conventional ANC instead minimizes inner-microphone signal power but attenuates both desired speech and noise.Source: §1 · p. 1
- Prior soft-constrained SSANC uses a scalar trade-off across frequencies, while frequency-dependent alternatives are typically block-based and may not meet SSANC’s low-delay needs.MVDR and MPDR beamformers also do not explicitly account for acoustic leakage or the secondary path.Source: §1 · p. 1
- LTASS and oracle weightings provide the largest intelligibility improvements and comparable speech-quality improvements at substantially lower distortion than SII and MIRS; oracle offers the best overall trade-off, while LTASS is comparable without clean desired speech.These are the paper’s reported perceptual comparisons.Source: §1 · p. 1
2. SIGNAL MODEL
The signal model represents the inner-microphone signal as acoustic leakage plus the secondary-source contribution filtered through the secondary path. It models outer-microphone inputs as desired speech and noise, and uses their signals with a leakage estimate in the control filter.
2. SIGNAL MODEL
- Each outer-microphone signal is decomposed into desired-speech and noise components.Source: §2 · p. 2
- The inner-microphone signal e(n) is the acoustic superposition of leakage p(n) and secondary-source signal y(n) filtered by the secondary path g.Source: §2 · p. 2
- The model assumes an available secondary-path estimate, and the analysis later assumes that estimate is exact, so the leakage estimate equals the leakage signal.Source: §2 · p. 2
- The hybrid SSANC architecture feeds outer-microphone signals and a leakage estimate to the control filter.The leakage estimate is extracted from the stacked input vector.Source: §2 · p. 2
- The secondary-source contribution at the inner microphone is represented through the secondary-path convolution applied to the control filter and stacked input.The stacked input includes the outer-microphone signal histories and leakage estimate.Source: §2 · p. 2
SPATIALLY SELECTIVE ACTIVE NOISE CONTROL
SSANC can enforce desired-speech preservation exactly or relax it with a scalar penalty that trades preservation against noise reduction. The proposed extension weights preservation error by frequency while retaining low-delay time-domain control, with several speech-based weighting filters compared.
3.1. Hard- and frequency-independent soft-constrained SSANC
- Hard-constrained SSANC imposes desired-speech preservation exactly, while soft-constrained SSANC penalizes its error using scalar μ to trade preservation against noise reduction.Increasing μ emphasizes preservation; μ = 0 reduces to conventional ANC, and μ → ∞ approaches the hard-constrained solution.Source: §3.1 · p. 2
3.2. Proposed frequency-weighted soft-constrained SSANC
- A frequency-weighting filter applied to preservation error varies its penalty across frequencies while maintaining sample-by-sample time-domain processing without block-based processing.The overall trade-off between desired-speech preservation and noise reduction remains controlled by μ.Source: §3.2 · pp. 2–3
3.3. Speech weighting filters
- The comparison includes uniform scalar, SII band-importance, MIRS speech-communication, LTASS speech-energy, and oracle desired-speech-PSD weightings.The oracle represents knowledge of the complete desired-speech spectral distribution; Fig. 2 shows normalized magnitude responses for an example signal.Source: §3.3 · p. 3, Figure 2 · p. 3
4. EVALUATION
The evaluation compares frequency weightings in a measured multi-speaker open-fitting-hearable setup. LTASS and oracle offer the strongest intelligibility results at lower distortion than SII and MIRS, with LTASS close to oracle overall.
4.1. Acoustic setup
- The setup used paired open-fitting hearables on a KEMAR simulator, measured room impulse responses, one desired speaker, and two interfering speakers; paths were assumed accurately known.The room reverberation time was approximately 330 ms, and the baseline inner-microphone leakage SNR was −5 dB.Source: §4.1 · p. 3
4.2. Evaluation metrics
- Evaluation covered noise reduction, broadband SNR improvement, desired-speech distortion, speech quality, and intelligibility, using SI-SDR, PESQ, and ESTOI for the latter measures.PESQ and ESTOI improvements were computed as metric differences between the controlled signal and the reference signal.Source: §4.2 · pp. 3–4
4.3. Results
- Conventional ANC achieved maximum noise reduction of 39.0 dB with strong desired-speech distortion, while hard-constrained SSANC reached SI-SDR 19.6 dB at noise reduction of 9.5 dB.Varying µ for scalar weighting traces the transition between these operating limits.Source: §4.3 · p. 4
- SII raised maximum ∆SNR from 16.7 dB to 18.5 dB and maximum ∆PESQ from 0.90 to 1.01 versus scalar weighting, but its ∆PESQ peak occurred at SI-SDR 1.0 dB versus 7.2 dB.MIRS achieved maximum ∆PESQ 0.99 at SI-SDR 3.0 dB and maximum ∆ESTOI 0.34.Source: §4.3 · p. 4
- LTASS (maximum ∆PESQ 0.93 at SI-SDR 10.1 dB; maximum ∆ESTOI 0.38) and oracle (0.92 at 16.7 dB; 0.39) had the largest intelligibility improvements at higher SI-SDR than SII and MIRS.Across leakage SNRs from −10 dB to 5 dB, LTASS and oracle maintained high maximum ∆PESQ at substantially higher SI-SDR, and the oracle had the best overall perceptual trade-off while LTASS was closely comparable.Source: Figure 4 · p. 4, Table 1 · p. 4, §4.3 · p. 4
- Across leakage SNRs from −10 dB to 5 dB, SII and MIRS had larger maximum ∆PESQ at low SI-SDR, whereas LTASS and oracle provided the largest ∆ESTOI improvements at higher SI-SDR.Table 1 reports maximum ∆PESQ and ∆ESTOI over µ with corresponding SI-SDR for each leakage SNR.Source: Table 1 · p. 4, §4.3 · p. 4
5. CONCLUSION
The paper generalizes soft-constrained SSANC with frequency-dependent speech-preservation weighting while retaining low-delay time-domain control. LTASS and oracle improve intelligibility at lower distortion than SII and MIRS, and LTASS approaches oracle performance without clean desired speech.
5. CONCLUSION
- Frequency-dependent weighting generalizes soft-constrained SSANC while maintaining low-delay time-domain active control.The weighting is applied to the desired-speech preservation error.Source: §5 · p. 4
- SII and MIRS produce larger maximum SNR and speech-quality improvements than LTASS and oracle, but at larger desired-speech distortion.Source: §5 · p. 4
- LTASS and oracle provide the largest intelligibility improvements with comparable speech-quality improvements at substantially lower distortion than SII and MIRS; LTASS closely matches oracle without clean desired speech.These trends remain consistent across the considered leakage SNRs.Source: §5 · p. 4