SPHERE: A Multi-Attribute No-Reference metric for Binaural Speech Quality And Spatial Fidelity
Organizations: Google LLC
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.
Figures & tables
| Perceptual Metric | Underlying Acoustic/Psychoacoustic Cues | Engineering & Product Significance in Wearables |
|---|---|---|
| Speech Quality (MOS-SIG) | Speech distortion, timbral coloration, phase mismatch | Evaluates the impact of voice processing on speaker naturalness and clarity |
| Noise Intrusiveness (MOS-BAK) | Signal-to-Noise Ratio (SNR), background noise levels | Measures the effectiveness of ambient noise suppression for user comfort |
| Overall Quality (MOS-OVRL) | Integrates speech quality, noise suppression, and spatial naturalness | Primary metric for product benchmarking and competitor comparison |
| Listening Effort | Intelligibility, cognitive fatigue, acoustic clarity | Critical KPI for conversation boost and hearing assistance features |
| Loudness Comfort | Volume appropriateness, dynamic range compression | Evaluates comfort of automatic volume leveling under varying environments |
| Sound Localization | Interaural Time/Level Differences (ITD/ILD), HRTFs | Ensures spatial rendering preserves accurate sound source direction |
| Perceptual Metric | DNSMOS [ 15 ] | Non-Siamese MEL | Siamese MEL | Non-Siamese STFT | Siamese STFT | Non-Siamese MEL + SF | Siamese MEL + SF |
|---|---|---|---|---|---|---|---|
| Overall Quality (MOS-OVRL) | 0.284 | 0.759 | 0.766 | 0.764 | 0.750 | 0.773 | 0.773 |
| Listening Effort | — | 0.776 | 0.781 | 0.781 | 0.769 | 0.784 | 0.783 |
| Background Noise (MOS-BAK) | 0.460 | 0.789 | 0.800 | 0.779 | 0.785 | 0.807 | 0.801 |
| Speech Quality (MOS-SIG) | 0.054 | 0.669 | 0.663 | 0.664 | 0.636 | 0.656 | 0.669 |
| Loudness Comfort | — | 0.755 | 0.745 | 0.747 | 0.736 | 0.759 | 0.756 |
| Sound Localization | — | 0.710 | 0.697 | 0.697 | 0.659 | 0.706 | 0.700 |
Speech Quality Assessment
Speech quality assessment measures how well speech preserves perceptual quality, intelligibility, pronunciation, naturalness, and speaker characteristics.
See all 75 papers