betaFeedback

WorksComputer Science

SPHERE: A Multi-Attribute No-Reference metric for Binaural Speech Quality And Spatial Fidelity

Abstract

In this paper, we introduce SPHERE (Spatial and Perceptual Human Evaluation and Rating Estimator), a multidimensional no-reference perceptual evaluation framework for binaural speech. SPHERE defines nine distinct perceptual and spatial attributes that characterize the immersive listening experience. We detail the crowdsourcing protocol and evaluation template used to collect ground-truth human ratings. To demonstrate the tractability of automating this framework, we evaluate six baseline configurations across single-branch and Siamese architectures utilizing Mel-spectrograms, complex STFTs, and pretrained spatial embeddings (PhaseCoder). Experimental results demonstrate that these baseline estimators achieve strong correlation with human subjective ratings, proving the feasibility of automated multidimensional spatial audio quality assessment.

INTRODUCTION

The proliferation of advanced hearables, smart glasses, and augmented reality (AR) devices has made spatial audio rendering and environmental noise suppression critical components of the modern acoustic experience [1, 2]. In these binaural systems, preserving precise interaural time and level differences (ITD and ILD) is essential for maintaining accurate sound localization and spatial immersion. However, aggressive non-linear signal processing—such as applying dual-channel neural noise suppression—frequently distorts these delicate spatial cues, leading to unnatural source wandering, localized artifacts, and increased listening effort [2, 3].

While traditional objective metrics like PESQ [4], POLQA [5], and ViSQOL [6] are highly effective for telephony speech, they are strictly single-channel and fail to capture spatial cues like sound localization and orientation. Recent deep-learning-based no-reference (NR) models, such as NISQA [7], DNSMOS [8], and SESQA [9], successfully predict multi-dimensional speech quality but remain blind to spatial degradation. Conversely, multichannel metrics such as PEAQ (ITU-R BS.1387) [10] and AMBIQUAL [11] evaluate perceptual similarity in ambisonics and spatial music. However, they require clean, full reference signals and do not assess speech-specific dimensions like listening effort, speech distortion, or neural voice enhancement artifacts.

To address these complexities, we developed a subjective testing template capturing nine distinct perceptual attributes. Built on established standards, it evaluates core quality and effort (ITU-T P.835 [12], P.810 [13]), spatial immersion (ITU-T P.830 [14]), and annoyance factors like artifacts and spatial wandering (ITU-T P.811/Supplement 29 [15]). To automate this evaluation, we introduce SPHERE, a no-reference objective framework predicting these nine attributes from stereo inputs. We evaluate six baseline configurations across single-branch and Siamese architectures utilizing Mel-spectrograms, complex STFTs, and pretrained spatial embeddings (PhaseCoder [16]). Experimental results demonstrate strong correlation with subjective human ratings, providing a foundational tool for evaluating binaural speech enhancement.

Subjective Evaluation Framework for Binaural Speech quality

Crowdsourcing Protocol and Hardware Pre-filtering

To collect reliable ground-truth perceptual ratings across large-scale binaural audio, we established a crowdsourced listening test protocol following the principles of ITU-T Recommendation P.808 [17]. Unlike monaural telephony evaluation, binaural spatial audio assessment is uniquely sensitive to listener hardware variability, as acoustic crosstalk from loudspeakers completely collapses interaural time and level differences (ITD/ILD).

To ensure measurement integrity, the protocol enforces a strict multi-tier pre-filtering procedure: (1) Hardware Pre-filter & Exclusion: Prior to accessing tests, raters declare their playback equipment category (wired over-ear, on-ear, in-ear earbuds, or built-in speakers). Laptop, phone, and external speakers are strictly disqualified. Hardware model details and Bluetooth/ANC state are logged to verify that spatial ratings are not contaminated by uncontrolled onboard DSP. (2) Binaural Channel Verification: To verify true stereo reproduction and correct ear orientation, raters must pass an acoustic screening test consisting of speech and tone tokens played exclusively in either the left or right ear, as well as binaural phase-cancellation checks. Raters failing this check are barred from the study. (3) Volume Calibration: Listeners adjust computer volume to a comfortable listening level during a standardized digit-in-noise sample and are instructed not to alter volume during evaluation. (4) Listener Qualification & Anti-Spam: Raters complete hearing screening questions (disqualifying individuals reporting hearing loss or prior involvement in speech coding) and spam/attention checks (transcribing 3 digits in background noise, and performing an audio math verification). Each test Human Intelligence Task (HIT) contains 10 evaluation clips, 1 hidden pristine gold anchor (rating < 4.0 disqualifies), and 1 degraded trap clip. Each clip receives a minimum of 5 screened ratings post-filtering (SEM ≈ 0.35 per clip); aggregating across conditions (> 100 clips per condition) reduces the standard error of the mean (SEM) to < 0.04, ensuring robust statistical significance for benchmarking.

Evaluation Template and Target Attributes

Ground-truth ratings are collected via a structured 4-part template: Part 1 (Core Quality & Effort) evaluates Speech Distortion (MOSSIG) and Background Intrusiveness (MOS-BAK) on 5-point ACR scales per ITU-T P.835 [12] (1: Very distorted/intrusive to 5: Not distorted/noticeable), Overall Quality (MOS-OVRL, 1: Bad to 5: Excellent) per P.800/P.835, and Listening Effort (1: Extremely high effort to 5: Very low effort / complete ease) adapted from ITU-T P.810 [13]. Part 2 (Spatial Immersion) assesses Loudness Comfort (P.800, P.830, P.840) to evaluate dynamic compression and Sound Localization (P.830) to evaluate directional orientation accuracy. Part 3 (Annoyance & Processing Artifacts) tracks Distracting Artifacts (clicks, pops, frame drops per P.830/P.835), Spatial Stability (unnatural source wandering per ITU-T P.811 / Supp. 29 [15]), and Background Noise Behavior (noise pumping or gating fluctuations per P.835). Part 4 (Conditional Low Score Justification) provides a structural validation mechanism. Triggered whenever a rater assigns MOS-OVRL ≤2\le2, the rater must identify the primary root cause (excessive background noise, speech distortion, loudness imbalance, or spatial instability). This cross-validates the inter-metric relationships between MOS-SIG, MOS-BAK, and MOS-OVRL, eliminating arbitrary scores and ensuring that the subjective MOS targets represent authentic, interpretable perceptual ground truth.

Datasets

To develop a robust no-reference perceptual quality metric for binaural speech, we compiled a large-scale database of binaural audio. This section describes the dataset compilation, the crowdsourcing perceptual labeling methodology, and the subsequent data augmentation and splitting protocols for model training.

Dataset Collection and Labeling

The core database consists of approximately 9,000 binaural audio clips of 10 to 20 seconds duration, designed to evaluate earbud playback and voice enhancement performance. The audio clips comprise: (1) Real-time captures recorded directly from commercial earbuds during active digital signal processing (DSP); (2) In-lab HATS recordings using a Head and Torso Simulator (HATS) in controlled acoustic chambers with varying speaker distance, azimuth angles, and background noise; and (3) Field recordings using mobile recorders to capture authentic environmental ambiance. The speech content includes voice signals processed by multiple configurations of neural voice enhancement algorithms alongside unprocessed baseline recordings. For each binaural clip, 5 human subjective ratings were collected via a crowdsourced listening test using the structured evaluation template detailed in Section 2. Ground-truth ratings were obtained across 9 distinct attributes of the perceptual and spatial experience, as summarized in Table 1.

Perceptual MetricUnderlying Acoustic/Psychoacoustic CuesEngineering & Product Significance in Wearables
Speech Quality (MOS-SIG)Speech distortion, timbral coloration, phase mismatchEvaluates the impact of voice processing on speaker naturalness and clarity
Noise Intrusiveness (MOS-BAK)Signal-to-Noise Ratio (SNR), background noise levelsMeasures the effectiveness of ambient noise suppression for user comfort
Overall Quality (MOS-OVRL)Integrates speech quality, noise suppression, and spatial naturalnessPrimary metric for product benchmarking and competitor comparison
Listening EffortIntelligibility, cognitive fatigue, acoustic clarityCritical KPI for conversation boost and hearing assistance features
Loudness ComfortVolume appropriateness, dynamic range compressionEvaluates comfort of automatic volume leveling under varying environments
Sound LocalizationInteraural Time/Level Differences (ITD/ILD), HRTFsEnsures spatial rendering preserves accurate sound source direction
Spatial StabilityAngular tracking stability, ITD/ILD drift over timeDetects unnatural wandering of voices due to head-tracking/dynamic-steering jitter
Distracting ArtifactsClicks, pops, frame drops, high-frequency digital glitchesMonitors low-latency DSP processing errors and codec-level distortions
Background Noise BehaviorNoise pumping, gating fluctuations, residual noise stabilityIdentifies distracting noise modulation artifacts caused by dynamic VADs

Table 1. The 9 target perceptual attributes, their underlying acoustic/psychoacoustic cues, and their engineering significance in spatial wearables (earbuds, smart glasses, and AR/VR headsets).

The distribution of the mean opinion scores (MOS) of 5 ratings across all 9 dimensions is illustrated in Figure 1, showing a comprehensive quality spectrum spanning the full range of 1.0 to 5.0. Figure 2 shows the correlation matrix between these metrics. We observe a very high correlation (0.93) between overall quality and listening effort, and a high correlation (0.81–0.82) between overall quality, spatial localization, and loudness comfort, demonstrating that spatial attributes are highly integrated into the overall human perceptual rating of binaural speech.

Distribution of human subjective Mean Opinion Scores across the 9 evaluated attributes of the binaural speech dataset

Figure 1. Distribution of human subjective Mean Opinion Scores (MOS) across the 9 evaluated attributes of the binaural speech dataset.

Pearson correlation matrix between the 9 subjective spatial audio metrics

Figure 2. Pearson correlation matrix between the 9 subjective spatial audio metrics, illustrating high redundancy between overall quality and listening effort, and moderate correlation between speech and noise quality.

Two-Stage Data Augmentation Pipeline

To train deep networks effectively and prevent overfitting, we implement a two-stage data augmentation pipeline using Apache Beam. The pipeline branches the source data into a pristine dataset and an augmented training corpus: (1) Sliding Window Chunking, where a 10-second window is slid across each 20-second source clip with a 2-second stride, extracting 6 chunks per file and propagating parent MOS labels under a weak supervision assumption. While overall quality remains stationary across 20 seconds, transient localized defects introduce label noise for unaffected chunks, explaining lower correlations on defect metrics; and (2) Perceptually Invariant Transformations, applying random time shifting (up to ±50\pm50 ms), gain adjustments (up to ±2.5\pm2.5 dB), and polarity inversion preserving MOS validity. The unmodified first chunk (0–10 s) forms the pristine dataset, while all sliding chunks and their invariant augmented variants form the augmented training corpus.

No-Leak Splitting Protocol

To ensure a rigorous evaluation and prevent data leakage across partitions, we enforce a strict speaker- and session-level splitting protocol. The partitioning is performed on the unique clean base audio file-names before any chunking or augmentation transformations are applied. This guarantees that all augmented variations (simulated rooms, spatial positions, gains, and window crops) of a given source audio file are mapped strictly to the same partition, preventing any overlap of speaker characteristics or environmental acoustic contents between the training and validation sets. The base clean filename is resolved for every record by stripping all augmentation and chunk suffixes. Shuffling and splitting (80% training, 20% validation) are performed directly on this list of unique base filenames. All subsequent augmented variations of a clean audio file are mapped strictly to the same partition. Prior to feature extraction, waveforms are normalized to [−1.0,1.0][-1.0, 1.0] by scaling the 16-bit integer representation by 32768.0. While the training set is fully augmented, we apply deterministic under-sampling to the validation split to keep the evaluation size manageable while preserving diversity. For each base file in the validation split, we keep the pristine chunk, and randomly select only 2 sliding window chunks (out of 5 possible) and 2 gain steps (out of 4).

The resulting split yields a training set of 433,396 records (1.394 TB stored in ArrayRecord format across 50 shards) and a multi-dimensional validation set of 12,554 records (∼\sim40 GB) categorized into standard, gain-augmented, and sliding-window variants.

Model Architectures

To establish the feasibility of predicting the 9 perceptual MOS metrics automatically, we evaluate two modeling paradigms: a single-branch (Non-Siamese) network and a two-branch (Siamese) weight-shared network. Both models employ a hybrid CNN-RNN architecture designed to capture spectral, temporal, and spatial cues from 10-second stereo audio segments.

Stereo Baseline Model (Non-Siamese)

The Non-Siamese baseline processes stereo audio by first extracting dual-channel features (either log-magnitude Mel-spectrograms or Short-Time Fourier Transform (STFT) magnitudes) from the left and right channels. These features are stacked along the channel dimension, yielding a tensor of shape [B,T,F,2][B,T,F,2], where BB is batch size, TT is time frames, and FF is frequency bins. This 2-channel representation is fed into a 2D CNN encoder consisting of five convolutional blocks with filter sizes (32, 32, 64, 64, 128) and batch normalization, followed by max-pooling layers.

The output is reshaped to [B,T,D][B,T,D] (D=1024D = 1024 for Mel). A temporal trunk of 2 stacked BiLSTM layers (128 units) captures bidirectional temporal context. Global Average Pooling yields a fixed-size representation of shape [B,256][B,256]. For multi-task regression, this is routed to 9 independent heads (64-unit Dense with ReLU, dropout 0.5, and 1-unit linear projection), totaling ∼\sim321k parameters.

Symmetric Two-Branch Siamese Architecture

Binaural perception relies on Interaural Time Differences (ITD) and Level Differences (ILD). To explore symmetric processing, we evaluate a Siamese architecture [18, 19]. As shown in Figure 3, left and right channels are processed by identical, weight-shared 2D CNN encoders (Shared θ\theta). For STFT, complex STFT features [B,T,F,2][B,T,F,2] stack real and imaginary parts.

Detailed architecture of the two-branch Siamese network

Figure 3. Detailed architecture of the two-branch Siamese network evaluated in SPHERE.

The CNN outputs are merged in a fusion layer hfused=[hL;hR;∣hL−hR∣;hL⊙hR]h_{\mathrm{fused}} = [h_L; h_R; |h_L - h_R|; h_L \odot h_R] and projected to match the baseline BiLSTM input dimension, maintaining comparable capacity (∼\sim330k parameters).

Spatial Feature Fusion (MEL + SF)

To evaluate self-supervised spatial embeddings, the MEL + SF configuration augments the Mel baseline with pretrained PhaseCoder [16] representations. PhaseCoder extracts a 256-D embedding capturing microphone-geometry agnostic interaural phase structures. This vector is z-score normalized, projected via a 64-unit Dense layer with SiLU, and concatenated with the 256-D pooled BiLSTM trunk ([B,320][B, 320]) before routing to the 9 prediction heads (∼\sim353k params).

Experimental Evaluation

Training Setup

Models are trained with Adam (learning rate 2×10−42 \times10^{-4}, batch size 128, TPU) with L2 regularization (10−410^{-4}). The objective is Weighted Mean Squared Error (WMSE) with equal attribute weights (wk=1.0,∀kw_k = 1.0, \forall k): L=(1/N)∑k=19wk(y^n,k−yn,k)2L = (1/N) \sum_{k=1}^{9} w_k(\hat{y}_{n,k} - y_{n,k})^2. Checkpoints are validated every 500 steps, retaining the model with maximum validation PCC.

Evaluation Metrics and Methodology

Models are benchmarked across five complementary statistical criteria: (1) Clip PCC (rpr_p): Pearson correlation coefficient evaluating linear alignment across all N=2,119N = 2{,}119 individual validation clips. (2) Clip SRCC (rsr_s): Spearman rank correlation coefficient assessing monotonic ranking accuracy invariant to non-linear score compression. (3) Absolute Error (RMSE, MAE): Root Mean Square Error and Mean Absolute Error quantifying absolute deviations on the 1.0–5.0 MOS scale. (4) System-Level PCC (rsysr_{\mathrm{sys}}): Evaluates algorithm benchmarking fidelity per ITU-T P.808/P.835. For each enhancement condition s∈Ss \in\mathcal{S} (N=6N = 6 systems, e.g., unprocessed audio, classic beamforming/DSP, and 4 neural voice enhancement baselines), clip ratings and predictions are averaged (yˉs,y^s\bar{y}_s, \hat{y}_s). System PCC is the Pearson correlation across condition means, filtering per-clip idiosyncratic acoustic variance. (5) Quality Tier PCC (rtierr_{\mathrm{tier}}): Evaluates global calibration linearity across 5 discrete subjective quality intervals ([1,2),(2,3),(3,3.5],(3.5,4],([1, 2), (2, 3), (3, 3.5], (3.5, 4], and (4,5](4, 5]) by averaging predictions and ratings per tier.

Clip-Level Correlation & Error: All configurations show robust clip-level tracking (Avg. PCC ≈0.70\approx0.70, SRCC ≈0.66\approx0.66, RMSE ≈0.73\approx0.73, MAE ≈0.58\approx0.58). Augmenting Mel spectrograms with PhaseCoder embeddings (MEL + SF) achieves the highest clip-level correlation (Avg. PCC = 0.706) and lowest error (Avg. RMSE = 0.726, MAE = 0.582 in Siamese MEL + SF), while improving rank correlation on Overall Quality (attribute-level SRCC = 0.708 vs. 0.687). Furthermore, Quality Tier PCC exceeds 0.99 across all models, confirming monotonic calibration across the degradation spectrum.

System-Level Ranking vs. Clip-Level Noise (Why Siamese?): While unconstrained 2D filters in Non-Siamese models show a marginal advantage on raw per-clip PCC (0.706 vs. 0.702), the Siamese architecture decisively outperforms Non-Siamese at the System Level (Avg. System PCC of 0.770 vs. 0.715). Weight-shared Siamese encoders enforce physical symmetry and interaural consistency (ITD/ILD), yielding dramatic gains on spatial attributes: Sound Localization jumps from 0.793 to 0.910, and Spatial Stability jumps from 0.786 to 0.867. Siamese models also excel on Background Noise (0.715 vs. 0.665) and Distracting Artifacts (0.822 vs. 0.768), confirming that symmetric inductive bias is essential for benchmarking spatial audio systems. Comparison with Monaural Baseline (DNSMOS): To quantify the need for native binaural modeling, we evaluated the open-source ITU-T P.835 DNSMOS predictor [8] (16 kHz) on the same validation set (N=2,119N = 2{,}119). Since DNSMOS accepts only single-channel inputs, we evaluated both time-domain downmixing (xmono=(xL+xR)/2x_{\mathrm{mono}} = (x_L + x_R)/2) and standard per-channel score averaging (y^=(f(xL)+f(xR))/2\hat{y} = (f(x_L) + f(x_R))/2). Acoustic downmixing introduces severe phase cancellation and comb-filtering due to inter-aural time differences (ITDs) and HRTFs (MOS-SIG rp=0.039r_p = 0.039, MOS-OVRL rp=0.256r_p = 0.256). While per-channel score averaging (reported in Table 2) avoids comb-filtering and improves correlation (MOS-OVRL rp=0.284r_p = 0.284, MOS-BAK rp=0.460r_p = 0.460, MOS-SIG rp=0.054r_p = 0.054), it still lags behind SPHERE by +0.347 to +0.615 PCC points because monaural models cannot capture binaural unmasking or interaural spatial cues and cannot predict listening effort or spatial attributes.

Perceptual MetricDNSMOS [8]Non-Siamese MELSiamese MELNon-Siamese STFTSiamese STFTNon-Siamese MEL + SFSiamese MEL + SF
Overall Quality (MOS-OVRL)0.2840.7590.7660.7640.7500.7730.773
Listening Effort—0.7760.7810.7810.7690.7840.783
Background Noise (MOS-BAK)0.4600.7890.8000.7790.7850.8070.801
Speech Quality (MOS-SIG)0.0540.6690.6630.6640.6360.6560.669
Loudness Comfort—0.7550.7450.7470.7360.7590.756
Sound Localization—0.7100.6970.6970.6590.7060.700
Spatial Stability—0.6370.6270.6440.6030.6430.636
Distracting Artifacts—0.6320.6070.6250.5730.6180.611
Background Noise Behavior—0.5930.5680.6040.5400.6110.589
Average Clip PCC (rpr_p ↑)—0.7020.6950.7000.6730.7060.702
Average Clip SRCC (rsr_s ↑)—0.6600.6520.6550.6300.6590.660
Average Clip RMSE (↓)—0.7440.7330.7420.7520.7340.726
Average Clip MAE (↓)—0.5990.5890.6000.6040.5910.582
Average System PCC (rsysr_{\mathrm{sys}} ↑)—0.7150.7700.7520.6880.7280.730
Average Quality Tier PCC (rtierr_{\mathrm{tier}} ↑)—0.9940.9940.9940.9920.9940.993

Table 2. Performance comparison across the 9 perceptual dimensions on the leak-free validation set (N=2,119N = 2{,}119 clips). Attribute-level Pearson Correlation Coefficients (Clip PCC, rpr_p) and 9-attribute model averages across all evaluation metrics are reported. For the monaural DNSMOS baseline [8], standard per-channel score averaging ((f(xL)+f(xR))/2f(x_L) + f(x_R))/2) is reported on its 3 supported P.835 dimensions (9-attribute averages are omitted as DNSMOS cannot predict 6 of the 9 dimensions). Best results in bold.

Conclusion

We introduced SPHERE, a 9-attribute subjective evaluation framework for binaural speech quality, listening effort, and spatial fidelity. Evaluated on a crowdsourced dataset with strict hardware pre-filtering and no-leak splitting, baseline neural architectures reliably predict all 9 attributes (up to 0.706 clip PCC, 0.770 system PCC, robust SRCC, and low RMSE/MAE). Crucially, system aggregation demonstrates that symmetric Siamese processing is vital for spatial fidelity (0.910 localization PCC), providing a foundational benchmark for next-generation hearables.

References

References

  1. [1]Jean-Marc Valin et al., “Audio processing for augmented reality and smart glasses,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.
  2. [2]Simon Doclo, Walter Kellermann, Shoji Makino, and Sven E Nordholm, “Binaural noise reduction algorithms for hearing aids,” IEEE Signal Processing Magazine, vol. 27, no. 1, pp. 11–30, 2010.
  3. [3]Marc René Schädler, Anna Warzybok, and Birger Kollmeier, “Binaural speech intelligibility benefits with directional microphones and noise reduction algorithms,” Trends in Hearing, vol. 24, pp. 2331216520938220, 2020.
  4. [4]A.W. Rix, J.G. Beerends, M.P. Hollier, and A.P. Hekstra, “Perceptual evaluation of speech quality (pesq)-a new method for assessing the quality of speech networks and codecs,” in 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), 2001, vol. 2, pp. 749–752 vol.2.
  5. [5]John G. Beerends, Christian Schmidmer, Jan Berger, Maarten Oome, Rajmund Ptasinski, and Cornelis Hoogendoorn, “Perceptual objective listening quality assessment (polqa), the third generation itu-t standard for end-to-end speech quality measurement part i–temporal alignment,” Journal of the Audio Engineering Society, vol. 61, no. 6, pp. 366–374, 2013.
  6. [6]Andrew Hines, Eoin Gillen, Damien Kelly, Jan Skoglund, Anil Kokaram, and Naomi Harte, “Visqualudio: An objective audio quality metric for low bitrate codecs,” in The Journal of the Acoustical Society of America. Acoustical Society of America, 2015, vol. 137, pp. 2298–2298.DOI
  7. [7]Gabriel Mittag, Babak Naderi, Assmaa Chehadi, and Sebastian Möller, “NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,” in Interspeech 2021. Aug 2021, ISCA.DOI
  8. [8]Chandan KA Reddy, Vishak Gopal, and Ross Cutler, “DNS-MOS P.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 886–890.
  9. [9]Joan Serrà, Jordi Pons, and Pascual Santiago, “Sesqa: Semi-supervised learning for speech quality assessment,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 381–385.arxiv.org/abs/2010.00368
  10. [10]ITU-R, “Method for objective measurements of perceived audio quality,” Recommendation BS.1387-2, International Telecommunication Union, Geneva, Switzerland, 2014.
  11. [11]Miroslaw Narbutt, Jan Skoglund, Andrew Allen, and Michael Chinen, “AMBIQUAL: Towards a quality metric for ambisonics,” in Proc. AES International Conference on Audio for Virtual and Augmented Reality, 2020.
  12. [12]ITU-T Recommendation P.835, “Subjective test methodology for evaluating speech communication systems that include noise suppression algorithm.,” 2003.
  13. [13]International Telecommunication Union, “P.810: Modulated noise reference unit (mnru),” Recommendation ITU-T P.810, International Telecommunication Union, 2023.
  14. [14]ITU-T, “Subjective performance assessment of telephone-band and wideband digital codecs,” Recommendation P.830, International Telecommunication Union, Geneva, Switzerland, Aug. 1996.
  15. [15]International Telecommunication Union, “ITU-T Recommendation P.811: Subjective test methodology for evaluating speech oriented stereo communication systems over headphones,” Recommendation P.811, International Telecommunication Union, Geneva, Switzerland, 2019.
  16. [16]Artem Dementyev, Wazeer Zulfikar, Sinan Hersek, Pascal Getreuer, Anurag Kumar, and Vivek Kumar, “PhaseCoder: Microphone geometry-agnostic spatial audio understanding for multimodal LLMs,” Proc. ICML, 2026.arxiv.org/abs/2601.21124
  17. [17]ITU-T Recommendation P.808, “Subjective evaluation of speech quality with a crowdsourcing approach.,” 2018.
  18. [18]Gabriel Mittag and Sebastian Möller, “Full-reference speech quality estimation with attentional siamese neural networks,” in Proc. ICASSP, 2020, pp. 346–350.
  19. [19]Pedro Morgado, Nuno Vasconcelos, Timothy Langlois, and Oliver Wang, “Self-supervised generation of spatial audio for 360° video,” in Proc. NeurIPS, 2018.arxiv.org/abs/1809.02587

Paper details

vlnc:W_swmp2575
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents