PaperlayerView exact revision on arXiv

Source-linked AI summary

UDSS-BWE: Uncertainty- and Decision-Science Inspired Swin BandWidth Extension

Tarikul Islam Tamiti, Sajid Fardin Dipto, David Vergano, Luke Baja-Ricketts, Anomadarshi Barua

arXiv:2609.36379v1cs.SD

TL;DR

Sparse, perceptually important high-frequency speech cues are difficult to infer from bandwidth-limited input, while average-case critics can dilute rare failures. UDSS-BWE combines a Swin-based dual-stream generator with decision-science and uncertainty-aware critics; it reports the best VCTK English NISQA-MOS and fewer parameters than AP-BWE. The authors report a slight training-time increase relative to AP-BWE.

  • Problem

    Sparse HF speech cues are weakly implied by narrowband observations, making their reconstruction ambiguous even though they affect naturalness, intelligibility, and downstream usability.

    Source: §1 · p. 1

  • Method

    UDSS-BWE combines a dual-stream Swin-1D lattice generator with five critics that address tail errors, HF over-boost, spectral criteria, epistemic uncertainty, and hard segments.

    Source: §1 · pp. 1–2

  • Results

    3.89x fewer parameters than AP-BWE (18.5M vs. 72M); UDSS-BWE achieves the best VCTK English NISQA-MOS and generalizes to MLS French.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The reported results support using decision-science and uncertainty theories as inductive biases for parameter-efficient BWE frameworks.

    Source: §1 · p. 2

    UDSS-BWE has a slight training-time increase relative to AP-BWE, mainly from added discriminator-side supervision and feature extraction.

    Source: §5 · p. 9

Abstract

from arXiv · show

Bandwidth extension (BWE) is fundamentally localized: the most perceptual distortions are not average-case distortions, but rare high-frequency (HF) transients that standard, risk-neutral objectives tend to smooth away. To close this gap, we seek solutions in the risk-sensitive and uncertainty-aware decision science rules and present UDSS-BWE, which introduces five decision-science and uncertainty-aware discriminators: CVaRD (does tail pooling to amplify HF artifacts), CCD (a primal-dual augmented Lagrangian to prevent HF overboost), MCUD (a learnable utility over spectral flatness/ centroid/ rolloff), EDD (captures epistemic uncertainty), and DROD (captures entropic KL-DRO aggregation). UDSS-BWE is also designed as a complex valued adversarial BWE framework that uses Swin-based generators, a lightweight dual-stream shifted-window backbone, to capture local and long-range structure efficiently, while learnable lattice coupling provides controlled cross-stream exchange. UDSS-BWE is optimized extensively and achieves better perceptual quality with 3.89x fewer parameters (72M vs.18.5M) over two English and French datasets under clean and noisy conditions. To the best of our knowledge, this work shows how multi disciplinary decision-science-inspired and uncertainty theories can be successfully used to design efficient discriminators for producing more nuanced audios, establishing a new baseline in the BWE task.

1 Introduction and Related Work

UDSS-BWE addresses the difficulty of reconstructing sparse, perceptually important high-frequency speech cues from ambiguous narrowband input. It combines a dual-stream Swin generator with decision-science and uncertainty-aware critics, and reports perceptual gains with fewer parameters than AP-BWE.

1 Introduction and Related Work

  • Sparse high-frequency cues such as fricatives and stop bursts are weakly implied by narrowband speech, making their timing and strength ambiguous despite their importance to naturalness and intelligibility.The passage notes that these cues have highly phoneme-dependent structure.Source: §1 · p. 1
  • Generic adversarial critics average errors, which can dilute rare HF failures, while sharper feedback can encourage implausible amplification or hiss-like artifacts.The paper identifies HF-sensitive supervision, training stability, and efficiency as remaining challenges.Source: §1 · p. 1
  • UDSS-BWE pairs a parameter-efficient dual-stream Swin-1D lattice generator with five critics that target tail errors, HF over-boost, spectral criteria, epistemic uncertainty, and hard segments.The magnitude and phase streams are refined in parallel with controlled cross-stream interaction.Source: §1 · pp. 1–2
  • 3.89x fewer parameters than AP-BWE (18.5M vs. 72M), with 2.66x MACs and FLOPs reduction; UDSS-BWE also achieves the best VCTK English NISQA-MOS and generalizes to MLS French.The paper also reports robustness in noisy conditions and subjective evaluation.Source: §1 · p. 2

2 UDSS-BWE Architecture

UDSS-BWE uses parallel magnitude and phase streams in a shifted-window Swin generator, then applies compact decision-science and uncertainty-aware critics to localized speech artifacts. Its critics emphasize rare tail errors, constrain persistent HF over-boost, evaluate interpretable spectral criteria, expose uncertainty, and emphasize difficult temporal segments.

2.1 Generator Design (see Fig. 1 & Alg. 6)

  • The generator processes magnitude and phase in parallel after the narrowband STFT, using shifted-window Swin blocks and learnable lattice coupling for controlled stream interaction.Windowed attention alternates offsets to propagate information across windows with O(TW) rather than O(T^2) complexity.Source: §2.1 · p. 2

2.2 Decision-Science Discriminators

  • Mean-pooled critics can dilute rare, perceptually harmful HF artifacts, motivating compact decision-science critics built with depthwise-separable 1D convolutions.The paper introduces CVaRD, CCD, and MCUD to encode interpretable HF artifacts.Source: §2.2 · p. 3

2.2.1 Conditional Value at Risk Dis. (CVaRD)

  • CVaRD averages the largest temporal activations with a global average summary, directing critic attention toward short, high-risk transients.Its tail fraction controls how many high-magnitude activations are selected.Source: §2.2.1 · p. 3

2.2.2 Chance-Constraint HF Discri. (CCD)

  • CCD reuses a differentiable HF-barrier trace in a primal-dual augmented-Lagrangian penalty to discourage persistent HF over-boost while preserving legitimate transients.Constraint enforcement activates progressively with GAN training through the adversarial warmup factor.Source: §2.2.2 · pp. 3–4

2.2.3 Multi-Criteria Utility Discri. (MCUD)

  • MCUD learns a convex utility over spectral flatness, normalized centroid, and 85% rolloff so one spectral improvement cannot conceal degradation in another.The criteria are normalized over time before their weighted utility trace is formed.Source: §2.2.3 · p. 4

2.3.1 Evidential Dirichlet Discri. (EDD)

  • EDD outputs Dirichlet evidence over real and fake classes, distinguishing weak evidence from strong evidence for the wrong class in ambiguous HF regions.The uncertainty-aware discriminator is built on a lightweight SepConv backbone.Source: §2.3.1 · p. 5, §2.3 · p. 5

2.3.2 Entropic KL-DRO Discri. (DROD)

  • DROD retains per-segment critic scores and aggregates hinge violations with an entropic-risk objective, emphasizing difficult temporal regions without unstable hard-max pooling.Increasing the risk parameter shifts the objective from average-case toward tail-sensitive supervision.Source: §2.3.2 · pp. 5–6, D and LDRO · p. 6

2.4 Loss Functions and Training Objectives

  • The generator objective combines reconstruction, feature-matching, and adversarial terms with EDD and DROD generator losses, the CCD penalty, and an optional uncertainty-weighted waveform L1 term.The adversarial warmup factor and weights govern the generator-side uncertainty losses.Source: §2.4 · p. 6

D + wEDD

The discriminator objective includes EDD and DROD terms, while MRAD and MRPD are adopted from prior work.

D + wEDD

  • The discriminator-side loss includes EDD and DROD losses.These are identified as discriminator-side terms in the training objective.Source: D + wEDD · p. 6
  • MRAD and MRPD are adopted from Lu et al. (2024).Source: D + wEDD · p. 6
  • The training objective distinguishes discriminator-side EDD and DROD losses from their generator-side terms.Source: D + wEDD · p. 6

3 Comprehensive Analysis

UDSS-BWE is evaluated across English and French speech, discriminator ablations, noisy and degraded-channel conditions, and perceptual and computational measures. It outperforms AP-BWE in reported perceptual comparisons while using fewer parameters and computation.

3.1 Metrics, Hyperparameters, and Datasets

  • Evaluation uses VCTK English and MLS French, totaling ∼1045 hours, and reports LSD, STOI, PESQ, SI-SDR, SI-SNR, NISQA-MOS, and WER.The datasets support multilingual and cross-speaker evaluation; MLS French lacks 48 kHz speech.Source: §3.1 · p. 6, Table 1 · p. 7

3.3 Comparative Analysis with Baselines

  • On VCTK 4→16 kHz, UDSS-BWE reports a 1.56× increase in NISQA-MOS versus unprocessed speech, the highest score among baselines, and improved perceptual reconstruction with 3.89× fewer parameters than AP-BWE (72M vs. 18.5M).Qualitatively, it reconstructs missing sibilants and brighter onsets more precisely than AP-BWE, approaching ground truth.Source: §3.3 · p. 7

3.2 Ablation Study

  • With MRAD and MRPD, N-MOS rises from ∼3.0 to 4.21–4.32, LSD falls to 0.98–0.99, and STOI stabilizes at 0.94; Row 10 reaches N-MOS 4.32.The ablation supports global structure recovery before local HF in-painting; EDD and DROD variants can achieve the best N-MOS despite lower SNR.Source: §3.2 · pp. 6–7, Table 2 · p. 7
  • Under the full discriminator suite, the AP-BWE generator collapses while the Swin-based generator remains high quality; proposed discriminators total ∼194.1k parameters.MRAD and MRPD together use 1.2M parameters.Source: §3.2 · p. 7

3.4 Evaluation under Noisy Conditions

  • UDSS-BWE consistently outperforms AP-BWE on averages across eight noise sources and five SNR levels, and is evaluated under channel conditions from sinc resampling alone to added filtering and packet loss.The channel study defines R0–R4, with codec mixtures introduced at R3 and telephone filtering plus packet loss added at R4.Source: §3.4 · p. 8, §3.5 · p. 8

3.6 Computational Complexity

  • UDSS-BWE uses 3.89× fewer parameters and 2.66× fewer MACs and FLOPs than AP-BWE, with comparable RTF and inference time and better perceptual quality.The reported complexity comparison associates these efficiencies with reduced memory use and suitability for real-time applications.Source: Table 5 · p. 8, §3.6 · p. 8

3.8 HF Analysis by ITU-Inspired Metrics; 3.9 Statistical Analysis of MOS Scores

  • Above 8 kHz, UDSS-BWE has lower error in 9 of 12 ITU-inspired comparisons, and paired Wilcoxon differences over AP-BWE are significant at utterance and speaker levels.The MOS statistical analysis covers 1000 utterances.Source: Table 6 · p. 8, §3.8 · p. 8, Table 8 · p. 9

3.7 Subjective Mean Opinion Score (MOS)

  • In a panel evaluation, UDSS-BWE scores MOS 4.25 versus AP-BWE’s 4.10; unprocessed audio scores 2.805.The panel comprised 10 persons.Source: §3.7 · p. 8

4 Conclusion

UDSS-BWE combines a parameter-efficient dual-stream generator with specialized discriminators for perceptually critical high-frequency failures. Experiments show improved wideband speech reconstruction across objective, perceptual, and downstream metrics, supporting decision-science and uncertainty theories as inductive biases for BWE.

4 Conclusion

  • UDSS-BWE combines a parameter-efficient dual-stream generator with specialized discriminator sets focused on perceptually critical high-frequency failures.The paper concludes that these decision-science and uncertainty theories provide effective inductive biases for parameter-efficient BWE frameworks.Source: §4 · p. 9
  • The authors report consistent improvements in wideband speech reconstruction across objective, perceptual, and downstream evaluation metrics.They identify ASR and TTS as relevant downstream tasks.Source: §4 · p. 9

5 Limitations

The authors report a slight training-time increase for UDSS-BWE relative to AP-BWE, attributed mainly to feature extraction in additional discriminator branches. The discriminator feature extractors differ in computational cost.

5 Limitations

  • The authors report a slight training-time increase relative to AP-BWE, stated as 17 minutes/epoch vs. 20 minutes/epoch.They attribute the added training cost mainly to feature extraction within the additional decision-science and uncertainty-aware critics.Source: §5 · p. 9
  • The Chance-Constraint and Multi-Criteria Utility feature extractors are identified as the most expensive decision-side components.Both require explicit time–frequency analysis and additional nonlinear processing before discrimination.Source: §5 · p. 9
  • The CVaR feature extractor adds noticeable overhead through tail-focused aggregation, while the uncertainty critics are comparatively lighter but still require extra forward and backward computation.The paper characterizes this as shifting computational burden to training for tail-aware, constraint-aware, and uncertainty-calibrated supervision.Source: §5 · p. 9

6 Potential Risks / Ethical Considerations

The paper identifies potential misuse of UDSS-BWE, including covert speech enhancement, speaker impersonation, and deepfake-style audio manipulation. Its appendices also describe the generator, training procedure, reconstruction losses, and phoneme-aware high-frequency priors.

6 Potential Risks / Ethical Considerations

  • Practical use should include clear disclosure, informed consent, and safeguards against misuse such as covert enhancement, speaker impersonation, or deepfake-style audio manipulation.The authors identify privacy, consent, and security concerns.Source: §6 · p. 9

A.1 Algorithms for the generator

  • The generator uses two lattice blocks with shared Swin1D weights to refine and couple log-magnitude and phase streams before predicting wideband magnitude and phase.The algorithm specifies a log-magnitude residual prediction and reconstructs phase from predicted components.Source: §A.1 · p. 11

A.3 Detailed Reconstruction Losses

  • The reconstruction-loss subsection defines magnitude, phase, complex-STFT, and STFT self-consistency terms for generator optimization.The passage identifies the target and generated complex STFTs and distinguishes magnitude and generated phase.Source: §A.3 · pp. 11–12

A.4.1 HF-band definition and phoneme-wise priors

  • The phoneme-wise priors distinguish weak high-frequency energy in sustained voiced speech from stronger, phoneme-dependent upper-band energy in voiceless fricatives.The cited findings support avoiding indiscriminate application of constraints suited to steady voiced segments to voiceless fricatives.Source: §A.4.1 · p. 12, G + LEDD,G + LDRO · p. 13

A.2 Algorithms for the training

  • The training algorithm combines reconstruction, adversarial, feature-matching, evidential, and entropic-risk terms, with a CCD chance constraint and dual-variable update.It uses warm-up weighting for adversarial loss and updates the generator and critics separately.Source: §A.4.1 · p. 13

G + LEDD,G + LDRO

  • The training procedure includes a CCD augmented-Lagrangian chance constraint and separate EDD and DROD generator terms.The algorithm also projects the CCD dual-variable update to its allowed interval.Source: G + LEDD,G + LDRO · p. 13

A.4.2 Decision-aware discriminator pressure on HF structure

UDSS-BWE’s discriminators target rare artifacts, phoneme-dependent high-frequency structure, and uncertain segments through complementary decision- and uncertainty-aware mechanisms. The appendix also details their computational costs, speech-specific rationale, and evaluation setup.

A.4.2 Decision-aware discriminator pressure on HF structure

  • CVaRD emphasizes the largest temporal responses, CCD penalizes excessive high-frequency excursions, MCUD assesses spectral descriptors, EDD exposes evidential uncertainty, and DROD aggregates segment violations through entropic risk.Together, these mechanisms target transient artifacts, overboost, phoneme-inconsistent spectra, ambiguous decisions, and heavy-tailed errors.Source: §A.4.2 · pp. 13–14

A.4.3 Details of Discriminator Design

  • The discriminator-cost analysis profiles five feature modules during training and notes that inference uses only the generator.Because standard MAC/FLOP tools do not fully account for parameter-free STFT criteria and top-k selection, the analysis adds analytic approximations for CVaRD, CCD, and MCUD.Source: §A.4.3 · p. 14

A.4.4 Analytic cost model

  • The analytic cost discussion characterizes CVaRD as lightweight tail pooling, CCD as an HF-ceiling constraint, and MCUD as a multi-objective combination of flatness, centroid, and rolloff.EDD models evidential uncertainty and DROD uses entropic-risk aggregation over segment scores.Source: §A.4.4 · pp. 14–15

A.4.5 Measured per-segment cost and overall interpretation

  • The dominant parameter-free training overhead comes from STFT-based feature computations in CCD and MCUD, rather than the decision logic itself.The reported comparison covers discriminator segments of 8,000 samples, or 0.5 s at 16 kHz.Source: §A.4.5 · p. 15

A.5.1 Why speech generation requires tail-aware supervision

  • Speech BWE errors can cluster in brief, high-impact events, which mean pooling or global spectral similarity may under-penalize and smooth over.Examples include fricative smearing, metallic chirps, transient HF spikes, and phonetic-boundary discontinuities.Source: §A.5.1 · pp. 15–16

A.5.2 Connection to phoneme-wise HF benchmarks

  • Phoneme-wise HF benchmarks motivate a vowel-oriented barrier against artificial broadband filling-in while preserving strong, natural fricative energy.EDD complements these mechanisms by moderating supervision when phonetic or acoustic ambiguity makes the real/fake boundary uncertain.Source: §A.5.2 · p. 16

A.5.3 Overall implication for UDSS-BWE

  • The overall design organizes adversarial training around tail risk, soft HF constraints, interpretable utility trade-offs, and calibrated epistemic uncertainty.The stated aim is to preserve microtransients and realistic phoneme-wise HF behavior without spurious amplification or brittle discriminator confidence.Source: §A.5.3 · p. 16

A.6 Data Preparation, Spectral Representation, and Inference

  • The evaluation uses VCTK English with a speaker-wise split and a 35-hour MLS French training version.For VCTK, 102 speakers are used for training and 8 reserved for testing; the reduced MLS subset was found inadequate for strong BWE performance.Source: §A.6 · p. 16

A.7 Evaluation Metrics

The appendix evaluates speech quality, intelligibility, and ASR compatibility, and reports computational and frequency-range analyses. Its ASR results show large gains over unprocessed narrowband input, while the frequency-range analysis finds stronger results as the reconstruction gap narrows.

A.7 Evaluation Metrics

  • The evaluation covers LSD, STOI, PESQ, SI-SDR, SI-SNR, NISQA-MOS, and WER, with WER measured using the open-source deepspeech-0.9.3 model.The metrics assess spectral deviation, intelligibility, perceptual quality, signal distortion, and ASR performance.Source: §A.7 · p. 17

A.8 WER Calculation for ASR Tasks

  • Under 4–16 kHz VCTK evaluation, character error rate falls from 68.59% to 10.19% and word error rate from 90.01% to 13.6% after UDSS-BWE enhancement.Word accuracy rises from 10.12% to 82.74%; similar gains are reported across evaluated bandwidth settings.Source: §A.8 · p. 17

A.9 Noisy Dataset Preparation and Detailed Noise Analysis

  • Noisy VCTK results are decomposed by SNR and noise type, while the preparation uses eight noise sources and five SNR levels.Noise and SNR are randomly selected per clean utterance, and a log records the selection and crop indices for reproducibility.Source: §A.9 · pp. 17–18

A.10 Details on Computational Complexity

  • UDSS-BWE has a GPU RTF of 0.0035–0.0033× versus AP-BWE’s 0.0023–0.0025×, which the authors describe as comparable and suitable for real-time downstream generation.Discriminator feature-extraction costs are reported separately, and their computations occur during training rather than inference.Source: §A.10 · pp. 18–19

A.11 Study for Different Frequency Ranges

  • Performance improves across all seven evaluation metrics as the gap between narrowband input and target bandwidth decreases.The 2–48 kHz reconstruction setting is the most challenging, while 24–48 kHz is the least challenging.Source: §A.11 · p. 19

A.12 ITU-Inspired Evidence of UDSS-BWE Improvement over AP-BWE

The section compares AP-BWE and UDSS-BWE using ITU-inspired sub-band diagnostics of temporal spectral continuity, loudness variation, and missing components. UDSS-BWE has lower error in most comparisons, including most generated high-frequency-band comparisons.

A.12 ITU-Inspired Evidence of UDSS-BWE Improvement over AP-BWE

  • UDSS-BWE has lower error in 12 of 18 comparisons and in 9 of 12 comparisons above 8 kHz, the generated high-frequency region.The three diagnostics are spectral-flux error, short-term loudness standard-deviation error, and missing-disturbance error; lower values indicate better alignment with the 48 kHz reference.Source: §A.12 · pp. 19–20, Table 18 · p. 19
  • Lower spectral-flux error indicates UDSS-BWE reconstructs high-frequency speech transitions with more reference-like temporal smoothness, because the diagnostic measures how the reconstructed high-frequency spectrum changes over time.Source: §A.12 · pp. 19–20
  • UDSS-BWE improves loudness-variation error in four of six bands and missing-disturbance error in four of six bands, including three of four generated HF bands.Reported missing-disturbance error reductions include 3.616 to 2.560 in 16–20 kHz; the metric penalizes reference components that are under-reconstructed.Source: §A.12 · pp. 19–20

A.13 Statistical Analysis of MOS Scores

The section tests MOS differences between UDSS-BWE and AP-BWE on a matched 1,000-utterance subset. UDSS-BWE has higher mean MOS and significantly higher scores in paired utterance- and speaker-level tests.

A.13 Statistical Analysis of MOS Scores

  • The MOS subset was sampled without replacement using a fixed random seed and an approximately balanced speaker allocation constrained by utterance availability.The subset contains 1,000 matched outputs from UDSS-BWE and AP-BWE.Source: §A.13 · p. 20
  • UDSS-BWE increases mean MOS over AP-BWE by 0.038 points on the 1,000-utterance subset, with higher MOS on 718 of 1,000 matched utterances.The utterance-level Wilcoxon test reports p = 1.94 × 10^-35; the speaker-level test reports p = 0.0313, with improvement in 6 of 7 speaker averages.Source: §A.13 · p. 20, Table 19 · p. 20, Table 20 · p. 20

A.14 PCM and Telephony-Codec Results

The section evaluates codec and telephony-channel conditions, documents model parameter breakdowns, and describes the training configuration and sensitivity analysis. Results show distinct trade-offs across channel augmentations, while the reported R4 setting specializes to degraded channels.

A.14 PCM and Telephony-Codec Results

  • PCM-8 augmentation substantially improves perceptual metrics over PCM-8 evaluation without that augmentation, while G.711-only augmentation does not protect against PCM-8 degradation.For the matched PCM-8 comparison, N-MOS rises from 1.8301 to 3.7645 while SI-SDR changes by only −0.0588 dB.Source: §A.14 · pp. 21–22
  • R4 performs poorly on ideal, PCM, and loss-free G.711 inputs despite its strong matched degraded-channel results, and its SI-SDR shifts from −19.7806 dB under ideal resampling to 7.8475 dB at 3% packet loss.The authors flag the unusual change for verification in a multi-seed analysis and note that current results favor condition-matched augmentation.Source: §A.14 · p. 22

A.15 Parameter Breakdown for Discriminators; A.16 Parameter Breakdown for Generators

  • The appendix provides layer-wise parameter breakdowns for CVaRD, CCD, MCUD, EDD, DROD, MRAD, MRPD, and the UDSS-BWE generator.The generator breakdown includes preprocessing, two LatticeBlock1D stages, and post-processing.Source: Table 23 · p. 23, Table 24 · p. 24, Table 25 · p. 25

A.17 Hyperparameters and Configuration

  • The training configuration lists the software and hardware environment and provides model and training hyperparameters with use cases and rationale.The reported environment includes Python 3.9.21, PyTorch 2.0.0+cu118, and CUDA Toolkit 11.8.0.Source: §A.17 · p. 27, Table 30 · p. 28, Table 31 · p. 28

A.18 Hyperparameter Sensitivity

  • Reference hyperparameters were chosen through preliminary development experiments rather than exhaustive grid or Bayesian search, and sensitivity tests are framed as local robustness analysis around the reported configuration.Sensitivity experiments use the 4–16 kHz VCTK setting and a 50-epoch training configuration; their anticipated qualitative effects are hypotheses, not empirical conclusions.Source: §A.18 · p. 27

A.19 Subjective Evaluation Details

A ten-volunteer pair-preference listening test compared UDSS-BWE with unprocessed and AP-BWE speech across three frequency-band conditions. Participants rated quality and chose a preferred version; UDSS-BWE was preferred most often.

A.19 Subjective Evaluation Details

  • Ten normal-hearing native English-speaking volunteers—four male and six female, average age 24—evaluated 30 sets using soundproof headsets after training.Participants joined voluntarily and without compensation.Source: §A.19 · pp. 29–31
  • Each 5–7-second VCTK clip had three randomly presented versions: noisy, AP-BWE-enhanced, and UDSS-BWE-processed speech; a clean reference was available but not evaluated.The clips covered 2–16 kHz, 12–48 kHz, and 24–48 kHz conditions.Source: §A.19 · p. 31
  • Participants rated each sample on a 1–5 MOS scale and selected their most preferred version; results were analyzed separately by participant sex and band configuration.Figure 6 shows the software interface used for the subjective evaluation.Source: §A.19 · p. 31, Figure 6 · p. 32
  • UDSS-BWE achieved ∼51% higher user preference than unprocessed speech and 3.6% higher preference than the AP-BWE baseline.Figure 5 presents the subjective test scores.Source: §A.19 · p. 31, Figure 5 · p. 31
Loading 2609.36379v1…