Source-linked AI summary
PRIME-ANC: Path-Ratio-Informed Modeling for Efficient Neural Filter Synthesis in Active Noise Control
Yaokun Huang, Chunyang Xu, Haowen Hua, Sen Lin, Shichao Hu, Mengyao Zhu
TL;DR
Listener acoustic changes require ANC filters to be redesigned for new paths, while existing noise-conditioned approaches can depend on system-specific resources or retraining. PRIME-ANC combines a regularized path-ratio base with learned path-dependent corrections to synthesize causal FIRs from calibrated paths. It improves held-out performance over the ratio-base FIR, and three Gauss–Newton updates approach direct WMMSE with lower measured design time, amplification, and control RMS.
Problem
Noise-conditioned filter-generation methods depend on system-specific filter resources or fixed-path training, limiting transfer to new acoustic environments.
Source: §1 · p. 1Method
PRIME-ANC combines calibrated path information, a regularized path-ratio magnitude base, and bounded path-dependent corrections to synthesize causal FIR filters.
Source: §1 · p. 1, §2.2 · p. 2Results
PRIME-ANC improves held-out noise reduction over the ratio-base FIR on both datasets; three Gauss–Newton updates approach direct WMMSE with lower core time, amplification, and control RMS.
Source: §4.1 · p. 4, §4.3 · p. 5Takeaways & Limitations
Calibrated paths let PRIME-ANC generate path-specific filters without query-specific optimization, and its FIRs can initialize Gauss–Newton refinement.
Source: §1 · p. 1
Abstract
from arXiv · showhide
Changes in listener acoustics require active noise control (ANC) filters to be redesigned for new acoustic paths. We introduce PRIME-ANC, a shared neural synthesizer that learns a bounded, path-dependent logmagnitude correction to a regularized path-ratio base. Minimum-phase reconstruction and truncation produce finite-impulse-response (FIR) filters; training optimizes their noise-control performance. Across ten random training/test splits per dataset, PRIME-ANC achieves average held-out one-third-octave reductions of 18.81 and 17.76 dB over 50 Hz-5 kHz on a ten-path dataset and a public earphone database, respectively. On original earphone measurements, it improves upon the path-ratio base by 7.89 dB. Ablation studies support the contributions of both the analytic base and the path-dependent correction. Given calibrated paths for a held-out listener condition, PRIME-ANC generates a path-specific FIR without iterative optimization. With three Gauss-Newton updates, it reaches 21.43 dB reduction, approaching direct weighted least-squares design while producing lower amplification and root-mean-square control output.
1. INTRODUCTION
Changing listener anatomy and earphone fit alter acoustic paths, making rapid ANC filter redesign important. PRIME-ANC uses calibrated paths in a shared generator to synthesize distinct causal FIR filters without retraining or query-specific optimization.
1. INTRODUCTION
- Listener anatomy and earphone fit alter acoustic paths, motivating rapid redesign of feedforward ANC filters.Existing approaches include online FxLMS/FxNLMS adaptation and fixed filters designed for target paths.Source: §1 · p. 1
- Existing noise-conditioned filter-generation methods remain tied to system-specific filter resources or fixed paths, often requiring resource preparation or retraining for a new acoustic environment.GFANC combines subfilters; E2E-CFG learns complete filters under fixed paths.Source: §1 · p. 1
- PRIME-ANC combines cross-path learning with calibrated path information to synthesize a distinct causal FIR for each query without retraining or query-specific optimization.The generator is trained on sparse support paths.Source: §1 · p. 1
- PRIME-ANC uses a regularized path-ratio magnitude base with bounded, path-dependent log-magnitude corrections.The approach also evaluates noise reduction, amplification, and gain after minimum-phase reconstruction and truncation.Source: §1 · p. 1
2. PRIME-ANC FILTER SYNTHESIS
PRIME-ANC trains a shared generator on calibrated path pairs and synthesizes query-specific causal FIR filters. It uses a regularized path-ratio magnitude base, bounded path-dependent corrections, and a realized-filter noise-control objective, with optional interpolation to extend sparse support coverage.
2.1. Acoustic paths and the filter-construction task
- A single network pass and deterministic realization produce a query-specific FIR without optimization.Estimated inputs guide synthesis and refinement; true support paths provide supervision and true query paths provide residuals.Source: §2.1 · p. 2
2.2. Path-ratio base and path-conditioned magnitude residual
- Regularization limits inversion near secondary-path spectral nulls, and the magnitude of the regularized path ratio forms the analytic base.Realization with zero learned correction yields the ratio-base FIR.Source: §2.2 · p. 2
- A shared CNN predicts bounded, path-dependent log-magnitude corrections, followed by minimum-phase reconstruction and truncation to a real, causal FIR.The loss is evaluated on the realized FIR because truncation can alter its magnitude and minimum-phase structure.Source: §2.2 · p. 2
2.3. Composite objective and support training
- Training penalizes worst-band amplification risk alongside target-band noise reduction.The risk term uses conditional value-at-risk on the worst 10% of bands from 50 Hz to 8 kHz.Source: §2.3 · pp. 2–3
- Optional paired interpolation uses one shared coefficient to interpolate adjacent support primary and secondary paths, while held-out conditions remain unchanged.Estimated inputs and true training paths share the interpolation coefficient.Source: §2.3 · p. 3
3. EXPERIMENTAL SETUP
The experiments evaluate PRIME-ANC on two acoustic-path datasets using listener-held-out splits, common noise-control scoring, and direct and iterative reference methods. They also measure query-specific design time under a defined runtime protocol.
3.1. Datasets, training splits, and evaluation
- The ten-path dataset contains ten path pairs, while PANDAR covers both ears of 23 participants; both include measured paths and specified path variants.Experiments use 48 kHz and 2048-tap filters, with eight DEMAND recordings plus white and pink noise.Source: §3.1 · p. 3
- Ten identity-based splits hold out 7 of 10 paths or 16 of 23 participants, with ears and variants kept together and fitting limited to support pairs.Each split adds 96 paired support mixtures; PRIME-ANC uses paired interpolation by default.Source: §3.1 · p. 3
- Noise reduction is scored over 50 Hz–5 kHz and amplification over 50 Hz–8 kHz; split means average held units after aggregation across noises, conditions, and seeds.The reported standard deviations describe ten paired, equally weighted split means, which are not independent trials.Source: §3.1 · p. 3
3.2. Direct synthesis and iterative references
- Direct references include zero-correction ratio-base FIRs, projected ratio filters, FOMAML initial filters, and a path-conditioned direct-tap baseline.Table 1 reports pre-adaptation filters without query updates; the path-conditioned baseline uses calibrated-path inputs and support-only paired interpolation.Source: §3.2 · p. 3, Table 1 · p. 4
- Iterative references include per-query Adam optimization and common-budget Gauss–Newton refinement of PRIME, analytic FIRs, FOMAML, and PANDAR WMMSE.FxLMS and FxNLMS use true secondary paths; the weighted FxLMS reference adapts for 40 seconds per path and noise.Source: §3.2 · p. 4, Figure 2 · p. 4
- PANDAR single-query core design time is measured after runtime initialization, excluding calibration, file access, model loading, and initialization.Calls are timed three times on an RTX 5090, except direct WMMSE on one CPU thread; condition medians are averaged within participants, seeds, and splits.Source: §3.2 · p. 4
4. RESULTS AND DISCUSSION
PRIME-ANC improves on the ratio-base filter in held-out direct synthesis, and ablations assess the base and path conditioning. Gauss–Newton refinement further approaches direct WMMSE performance with lower measured cost and output levels.
4.1. Direct synthesis on held-out acoustic conditions
- 2.857/7.663 dB NR gains over the ratio base on ten-path/PANDAR came with positive seed-averaged gains in 9/10 and 10/10 splits.The gains span much of the lower and middle target bands; mean AMP is 0.770/0.134 dB.Source: §4.1 · p. 4
- On original PANDAR measurements, PRIME reaches 18.576 dB NR versus 10.686 dB for the ratio base, and 16.832 dB without interpolation.With interpolation, modified conditions gain 7.588 dB; paired interpolation adds 2.083/1.539 dB on ten-path/PANDAR.Source: §4.1 · p. 4
4.2. What the ratio base and conditioning contribute
- 3.93/3.96 dB from the base with MP/phase-corrected outputs across all ten splits; phase correction adds < 0.04 dB, while path-independent correction and phase shuffling lose 5.246/5.162 dB NR.The ablation compares PANDAR construction and conditioning variants; the unbounded variant improves PANDAR but degrades ten-path NR.Source: §4.2 · p. 5, Table 2 · p. 5
4.3. Refinement and computational efficiency
- At comparable core times of about 2.1 ms, three PRIME GN updates achieve 2.14 dB higher NR than four ratio-base GN updates; PRIME+GN(3) approaches direct WMMSE with lower core time, AMP, and control RMS.The common-update comparison tests initialization under identical query estimates and objective.Source: Table 3 · p. 5, Figure 3 · p. 5, §4.3 · p. 5
5. CONCLUSION
PRIME-ANC synthesizes causal FIR filters from calibrated paths, combining an analytic path-ratio base with learned path-dependent corrections. It improves over the ratio-base FIR on held-out paths and listener conditions; three Gauss–Newton updates approach direct WMMSE on PANDAR with lower design time, amplification, and control RMS.
5. CONCLUSION
- On PANDAR, three Gauss–Newton updates approach direct WMMSE in noise reduction with lower measured core design time, amplification, and control RMS.The conclusion presents the updates as a refinement of PRIME-ANC's synthesized filters.Source: §5 · p. 5
- PRIME-ANC combines an analytic path-ratio base with learned path-dependent log-magnitude corrections to synthesize causal FIRs from calibrated paths.The framework is described as enabling rapid, listener-specific filter generation with lower design overhead and robust noise-control performance.Source: §5 · pp. 5–6
- Experiments show gains over the ratio-base FIR on held-out paths and listener conditions, while ablations support both the analytic base and path conditioning.Source: §5 · p. 5