Paper deep dive
Multichannel Audio Quality Assessment: Extending Pretrained Perceptual Models to Spatial Audio
Gouthaman KV, Shiv Gehlot, Vishnu Raj, Lars Villemoes, Arijit Biswas
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Accurate perceptual quality assessment is essential for evaluating and optimizing spatial audio, where perceived quality depends on both signal fidelity and inter-channel spatial relationships. However, subjective evaluation is costly, while existing perceptual models are often trained for limited channel configurations and cannot be directly applied to higher-channel-count audio. This raises the question: how can pretrained perceptual knowledge be effectively reused for multichannel spatial audio? Using 5.1-channel audio, we study four levels of multichannel integration: signal, prediction, latent, and feature and propose two learned approaches: latent-level aggregation of spatial-group representations and the feature-level Feature-Band Group Attention (FGAtt), which adaptively fuses spatial groups at the feature level before perceptual processing. Across five 5.1-channel test sets, FGAtt achieves the strongest over- all performance, demonstrating the effectiveness of feature-level adaptation for reusing pretrained perceptual knowledge
Tags
Links
- Source: https://arxiv.org/abs/2609.37116v1
- Canonical: https://arxiv.org/abs/2609.37116v1
Trouble viewing inline? Open PDF directly →
Full Text
32,710 characters extracted from source content.
Expand or collapse full text
Multichannel Audio Quality Assessment: Extending Pretrained Perceptual Models to Spatial Audio Gouthaman KV Shiv Gehlot Vishnu Raj Lars Villemoes Arijit Biswas Abstract Accurate perceptual quality assessment is essential for evaluating and optimizing spatial audio, where perceived quality depends on both signal fidelity and inter-channel spatial relationships. However, subjective evaluation is costly, while existing perceptual models are often trained for limited channel configurations and cannot be directly applied to higher-channel-count audio. This raises the question: how can pretrained perceptual knowledge be effectively reused for multichannel spatial audio? Using 5.1-channel audio, we study four levels of multichannel integration: signal, prediction, latent, and feature and propose two learned approaches: latent-level aggregation of spatial-group representations and the feature-level Feature-Band Group Attention (FGAtt), which adaptively fuses spatial groups at the feature level before perceptual processing. Across five 5.1-channel test sets, FGAtt achieves the strongest overall performance, demonstrating the effectiveness of feature-level adaptation for reusing pretrained perceptual knowledge. Index Terms: Spatial audio, multichannel audio quality metrics, perceptual audio quality †address: Dolby Laboratories 1 Introduction Objective perceptual audio quality assessment is essential to audio coding, enhancement, restoration, and spatial audio systems. Although subjective tests such as MUSHRA [11] directly measure perceived quality, their cost and limited scalability motivate objective metrics. Conventional metrics, including PEAQ [18] and ViSQOL [8], use predefined perceptual representations and signal-comparison procedures [10, 7], whereas learning-based approaches such as InSE-NET [3] and the Generative Machine Listener (GML) [12, 16] learn representations from subjective quality data and agree more closely with human judgments than conventional signal-level metrics. Thus, learned perceptual representations offer a powerful basis for objective quality prediction. The growing adoption of multichannel and spatial audio challenges perceptual models designed for fixed channel configurations. For instance, GMLv2 [16], a state-of-the-art learning-based perceptual model, supports mono, stereo, and binaural audio but cannot be directly applied to higher-channel formats such as 5.1 and 7.1.4. A straightforward solution is to reduce multichannel signals to a mono or two-channel representation through downmixing or binauralization [4, 13, 17, 5]. However, such signal-level transformation can lose perceptually relevant channel-specific information and inter-channel relationships. Another approach is prediction-level late fusion, where individual channels are processed independently and their predictions are aggregated. While this avoids explicit signal reduction, it cannot model inter-channel interactions and requires additional computation for each channel. Alternatively, a standalone multichannel model can learn these interactions directly, but requires huge amount of format-specific subjective data and training and does not directly leverage perceptual knowledge from existing models such as GMLv2. As spatial formats proliferate, training separate models becomes increasingly expensive and repeatedly relearns perceptual information, motivating a scalable approach that reuses pretrained perceptual knowledge while incorporating higher-channel audio. Consequently, this work examines how a mono/two-channel perceptual model can be adapted to higher-channel formats while preserving its perceptual knowledge. We use GMLv2 as the backbone, which achieves strong quality-prediction performance for mono, stereo, and binaural audio [16], and study its extension to 5.1 audio. Specifically, we compare four ways of incorporating multichannel information into its processing pipeline: signal-level transformation through downmixing or binauralization, prediction-level aggregation across channels or spatial groups, latent-level aggregation of spatial-group representations after perceptual processing, and feature-level fusion before perceptual processing. Within this framework, we propose two learned approaches that retain the pretrained GMLv2 backbone: latent-level aggregation of spatial-group perceptual representations and Feature-Band Group Attention (FGAtt), which adaptively fuses multichannel Gammatone features using input-dependent, feature-band-specific spatial-group weights before perceptual processing. We evaluate the proposed strategies on five 5.1-channel test sets and show that the feature-level adaptation method, FGAtt provides a more effective approach to reusing pretrained perceptual knowledge than signal-domain reduction, prediction-level aggregation, and latent-level aggregation. Overall, our findings highlight the importance of both the level and granularity of multichannel integration when extending pretrained mono/two-channel perceptual models to spatial audio. The results demonstrate a promising path for scaling pretrained perceptual knowledge to higher-channel-count audio while keeping the perceptual backbone frozen. 2 Methodology 2.1 Problem Formulation We aim to extend GMLv2 [16], a reference-based perceptual model trained on a large collection of subjective evaluations of coded mono, stereo, and binaural audio. GMLv2 achieves stronger correlation with human judgments than conventional metrics such as ViSQOL and PEAQ, but supports inputs of at most two channels. For left and right channels L and R, GMLv2 forms a four-component representation [L,R,(L+R)/2,(L−R)/2][L,\,R,\,(L+R)/2,\,(L-R)/2]. Gammatone features are extracted from these components and concatenated along the component dimension for each reference–degraded pair. A neural network then predicts the parameters of a Beta distribution, which are mapped to the MUSHRA scale to obtain the quality score. To extend GMLv2 to 5.1-channel audio, let rx^r and dx^d denote the reference and degraded multichannel audio signals, respectively. We define a multichannel integration operator Aθ(l)A_θ^(l) at one of four levels, l∈signal,prediction,latent,featurel∈\signal,prediction,latent,feature\, and formulate the quality prediction as: q^=ℱϕ(r,d,Aθ(l)), q=F_φ (x^r,x^d;A_θ^(l) ), -5.69054pt (1) where ℱϕF_φ denotes the pretrained GMLv2 backbone with parameters ϕφ kept frozen, q q is the predicted perceptual quality score, and Aθ(l)A_θ^(l) denotes the corresponding fixed or learned multichannel integration strategy. Depending on l, multichannel information is incorporated before or after perceptual processing, providing a common framework for comparing the four strategies while retaining the pretrained GMLv2 backbone. 2.2 Multichannel Adaptation We investigate four strategies for extending pretrained GMLv2 to 5.1-channel audio, categorized by the level at which multichannel information is integrated: signal, prediction, latent, or feature. Signal- and prediction-level strategies directly reuse the pretrained model, whereas latent- and feature-level strategies learn the integration within its processing pipeline. For all methods, we exclude LFE from the 5.1-channel input L,R,C,LFE,Ls,RsL,R,C,LFE,L_s,R_s to avoid false penalties when codecs remove reference energy above the standard 120-Hz LFE cutoff, which subwoofer bass management would normally filter during playback. At the signal level, ITU downmixing [9] or binauralization [4] transforms the 5.1-channel signal into a representation supported by GMLv2. While simple, these transformations use fixed signal-domain rules that may discard perceptually relevant spatial information, particularly when degradation characteristics vary across channels. Designing suitable fixed transformations also becomes increasingly difficult as the channel count grows, limiting their applicability to richer spatial formats. At the prediction level, GMLv2 processes the 5.1 channels either individually as mono inputs or in two-channel groups, consistent with its support for up to two channels. The resulting quality predictions are averaged to obtain the final score. Since predictions are combined only after independent processing, this approach cannot model inter-channel interactions or learn different contributions across inputs. Moreover, independent backbone evaluations increase computational cost with the number of channels or groups. The group-based variant uses the spatial grouping defined below. These limitations motivate learned integration mechanisms that adapt to the input. At the latent level, GMLv2 processes the groups independently, and learned attention weights combine their perceptual representations before quality prediction. At the feature level, group features are fused before perceptual processing, enabling the frozen backbone to jointly process multichannel information. In this regard, we propose Feature-Band Group Attention (FGAtt), which learns input-dependent, feature-band-specific weights over the groups. The groups and their common multichannel representation are defined next, followed by the respective latent- and feature-level integration strategies. Multichannel Feature Representation: The group-level prediction-, latent-, and feature-level approaches share a common spatial grouping consisting of the front pair, the surround pair, and the center channel: =LR,LsRs,CG=\LR,LsRs,C\. For each paired group (x1,x2)(x_1,x_2), we construct the four-component representation expected by GMLv2 [x1,x2,(x1+x2)/2,(x1−x2)/2]. [x_1,x_2,(x_1+x_2)/2,(x_1-x_2)/2 ]. The single center signal is duplicated according to the backbone’s mono-input procedure [16]. Independently extracting Gammatone features from the four components yields (g)∈ℝ4×F×DX^(g) ^4× F× D, g∈g , where F=32F=32 is the number of Gammatone bands and D is the temporal dimension. Thus, each spatial group is mapped to the pretrained backbone’s input space while preserving its spatial identity. Learned Latent-Level Aggregation: The first learned strategy introduced in this work fuses the spatial groups immediately before the GMLv2 prediction head. Each group g∈g is processed independently by the pretrained backbone, producing (g)∈ℝDzz^(g) ^D_z. Each latent representation encodes quality-relevant information from its reference–degraded pair using the perceptual representation learned by GMLv2. To learn the relative contribution of each group, we project its latent representation and compute an attention score: (g) ^(g) =tanh(p(g)+p),eg = (W_pz^(g)+b_p ),\ e_g =(g)⊤,∈ℝDh, =h^(g) v,\ \ v ^D_h, (2) where pW_p, pb_p, and the context vector v are learnable. The scores are normalized across groups and used to form a weighted combination of the latent representations: αg=exp(eg)∑g′exp(eg′),agg=∑gαg(g). _g= (e_g) _g (e_g ), 18.49988ptz_agg= _g _gz^(g). -5.69054pt (3) The frozen GMLv2 prediction head then maps aggz_agg to the final quality score. This approach preserves GMLv2’s perceptual processing within each spatial group while learning how their representations should be combined, providing a learned alternative to prediction-level aggregation. However, it requires a separate backbone inference for every spatial group, causing computational cost to grow with the number of groups. Feature-Band Group Attention (FGAtt): Latent-level aggregation combines spatial groups only after independent perceptual processing, preventing the backbone from modeling cross-group interactions. To enable joint processing, we instead fuse the groups at the feature level and pass the resulting representation through the frozen GMLv2 backbone. Since the perceptual contribution of each group may vary with both the input and Gammatone band, fixed fusion weights are insufficient. We therefore propose Feature-Band Group Attention (FGAtt), which learns input-dependent, band-specific weights over spatial groups. FGAtt follows three principles. First, temporal information is pooled only for attention estimation, leaving the temporally resolved features intact for fusion. Second, group importance is inferred jointly from the four GML-compatible components, preserving the spatial relationships encoded by the pretrained backbone. Third, a shared nonlinear mapping and cross-group normalization provide attention weights that adapt to both the input and the Gammatone feature band. Let g∈g denote a spatial group with (g)∈ℝ4×F×DX^(g) ^4× F× D representing its Gammatone features. We first pool the temporal dimension: (g)=1D∑d=1D(g):,:,d∈ℝ4×F,S^(g)= 1D _d=1^DX^(g)_:,:,d ^4× F, -5.69054pt (4) retaining the four-component and feature-band dimensions. At each band f, (g)f=(g):,f∈ℝ4s^(g)_f=S^(g)_:,f ^4 jointly describes the activity of the four components. A shared Multi-layer Perceptron (MLP) maps this vector to a scalar attention score, ef(g)=MLP(f(g))e^(g)_f=MLP\! (s^(g)_f ). Sharing the MLP across groups and bands keeps the adaptation module lightweight while providing a common scoring function. Because the scores are computed from the input features, they are input-dependent rather than fixed. The scores are then normalized across groups at each feature band: αf(g)=exp(ef(g))∑g′∈exp(ef(g′)),∑g∈αf(g)=1.α^(g)_f= \! (e^(g)_f ) _g \! (e^(g )_f ), 18.49988pt _g α^(g)_f=1. (5) The resulting weights determine the relative contribution of each spatial group at each feature band. They are applied to the original temporally resolved features: FGAtt:,f,d=∑g∈α(g)f(g):,f,d.X^FGAtt_:,f,d= _g α^(g)_fX^(g)_:,f,d. -5.69054pt (6) Thus, temporal pooling (4) affects only attention estimation, while the original temporal information is preserved during fusion. The same scalar weight is applied to all four components within a group and band, adjusting the group’s overall contribution while preserving its internal GML-compatible structure. FGAtt is applied independently to the reference and degraded representations, with attention weights determined by their respective features. The fused representations are then concatenated along the component dimension and passed to a single frozen GMLv2 backbone and prediction head. Unlike latent-level aggregation, which requires a separate backbone inference for each spatial group, FGAtt enables joint processing of all groups in a single inference. It therefore provides a lightweight interface between higher-channel spatial audio and a pretrained lower-channel model while preserving the expected feature structure and temporal information. Although evaluated on 5.1 audio, the same formulation can be extended to other spatial formats by defining appropriate channel groups and their corresponding group-level representations. 3 Experiments and Results 3.1 Experimental Setup Dataset: Collecting subjective quality data for multichannel audio is more involved than conventional low-channel evaluation, as perceived quality depends on both signal fidelity and spatial characteristics. As part of this study, we undertook a dedicated data collection effort to obtain subjective 5.1-channel quality data for training and evaluating the proposed learned adaptation strategies. Training: The training corpus contains approximately 26 hours of 5.1 audio and 21K samples with subjective scores from 75 MUSHRA tests conducted over several years of multichannel codec research and development. It covers multiple generations of coding technologies: Dolby Digital, HE-AAC, Dolby Digital Plus, and AC-4, across diverse bitrates and perceptual quality levels. Evaluation Sets: We evaluate the approaches on five independent subjective listening-test datasets (T1–T5). Three of them (T1–T3) are derived from the MPEG Surround verification tests [6], each containing ten 5.1-channel excerpts covering ambient, jazz, modern, orchestral, and popular music. T1, T2, and T3 were evaluated by 39, 35, and 18 listeners, respectively. T1 covers DVB-oriented coding scenarios, including Layer-2/MPEG Surround and Dolby Pro Logic I at 256 kb/s, as well as HE-AAC-based multichannel systems at 64 and 160 kb/s. T2 targets music distribution applications and compares AAC-based MPEG Surround, enhanced matrix decoding, and Dolby Pro Logic I systems operating at approximately 192 kb/s. T3 compares high-quality and low-power MPEG Surround decoding modes. The remaining two test sets (T4–T5) are internal listening tests constructed from fourteen immersive 7.1.4 excerpts by extracting the corresponding 5.1 channels. T4, evaluated by 14 listeners, includes Opus (192 kb/s), AAC (256 kb/s), HE-AAC (192 kb/s), MPEG Surround-based systems (192 kb/s), Dolby Digital Plus (192 kb/s), and Dolby Digital (384 kb/s). T5, evaluated by 13 listeners, includes Opus (160 kb/s), HE-AAC (128 and 160 kb/s), and MPEG Surround-based systems (96 and 128 kb/s). Together, the five evaluation sets comprise 428 samples and span diverse codec families, bitrate operating points, content types, and subjective quality levels, providing a broad basis for multichannel audio quality evaluation. Implementation: For signal-level experiments, we use ITU downmixing [9] and the Binamix binauralizer [2]. The learned approaches, latent-level aggregation and FGAtt, are trained on the aforementioned training set using the objective of the original GMLv2 training [16]. The GMLv2 backbone remains frozen, while the aggregation modules are optimized with AdamW [14] using a learning rate of 0.0010.001. All methods are implemented in PyTorch and trained on two NVIDIA A10 GPUs. 3.2 Results and Discussion Table 1 compares the evaluated methods across five test sets using Pearson correlation coefficient (PCC) and Spearman correlation coefficient (SCC) between predicted and subjective quality scores. The LFE channel is excluded throughout. The GMLv2-based methods cover prediction-, signal-, latent-, and feature-level integration, while Zimtohrli [1], PEAQ [18], and ViSQOL [8] provide reference baselines based on established perceptual metrics with mono or two-channel inputs. Zimtohrli [1] achieves an overall PCC/SCC of 0.7763/0.68720.7763/0.6872, with performance ranging from 0.7228/0.59070.7228/0.5907 to 0.9181/0.91380.9181/0.9138 across the five evaluation sets. As Zimtohrli is designed for monaural audio, its multichannel extension independently evaluates each non-LFE channel and averages the resulting predictions. While this provides a straightforward extension to multichannel audio, it requires a separate metric evaluation for each channel, with computational cost increasing linearly with the number of channels. Moreover, independent channel-wise predictions do not explicitly capture inter-channel relationships that contribute to perceived spatial audio quality. For PEAQ [18], ITU Downmix + PEAQ achieves an overall PCC/SCC of 0.4667/0.59440.4667/0.5944, while Binauralizer + PEAQ achieves 0.4051/0.51120.4051/0.5112. Both configurations perform substantially below their corresponding ViSQOL variants: ITU Downmix + ViSQOL achieves 0.8573/0.74550.8573/0.7455, compared with 0.8395/0.80660.8395/0.8066 for Binauralizer + ViSQOL. While ITU downmixing yields higher PCC with ViSQOL, binauralization achieves higher SCC and consistently outperforms ITU downmixing in SCC across all five evaluation sets. These results highlight that both the underlying perceptual metric and the choice of multichannel-to-two-channel transformation substantially affect multichannel quality prediction. Table 1: Perceptual quality prediction performance on five 5.1-channel test sets. LFE is discarded. GMLv2 is frozen. Zimtohrli evaluates each channel independently, with predictions averaged across channels. T1 T2 T3 T4 T5 Overall Method PCC SCC PCC SCC PCC SCC PCC SCC PCC SCC PCC SCC Zimtohrli [1] 0.7228 0.5907 0.9181 0.9138 0.8210 0.6948 0.8365 0.7245 0.8255 0.8129 0.7763 0.6872 ITU Downmix [9] + PEAQ [15] 0.5846 0.7686 0.7948 0.8753 0.5199 0.8200 0.3814 0.5376 0.3023 0.2182 0.4667 0.5944 Binauralizer [2] + PEAQ [15] 0.5684 0.7483 0.5970 0.6086 0.5216 0.8092 0.2969 0.4244 0.2187 0.0953 0.4051 0.5112 ITU Downmix [9] + ViSQOL [7] 0.8419 0.7428 0.8739 0.8643 0.8934 0.7480 0.8682 0.7720 0.8646 0.7125 0.8573 0.7455 Binauralizer [2] + ViSQOL [7] 0.8477 0.8791 0.8895 0.9204 0.8578 0.8798 0.8507 0.7810 0.8378 0.7549 0.8395 0.8066 Mono Pred-level + GMLv2 0.8305 0.8320 0.8743 0.8651 0.8565 0.9207 0.7942 0.5653 0.8146 0.4790 0.7464 0.5249 Group Pred-level + GMLv2 0.8759 0.8635 0.8921 0.8663 0.8846 0.9171 0.8279 0.7959 0.7808 0.7605 0.8131 0.7731 ITU Downmix + GMLv2 0.9247 0.8759 0.9213 0.8916 0.9488 0.8893 0.8493 0.7844 0.7976 0.6891 0.8623 0.7839 Binauralizer + GMLv2 0.8630 0.8498 0.8838 0.8503 0.8596 0.8716 0.8096 0.7626 0.7725 0.6947 0.8045 0.7564 Latent Agg + GMLv2 0.8578 0.8414 0.8747 0.8536 0.9086 0.8740 0.9093 0.7759 0.9350 0.8364 0.8875 0.7995 FGAtt + GMLv2 0.9001 0.8541 0.9207 0.8880 0.9106 0.8809 0.9142 0.7889 0.9415 0.8616 0.9063 0.8303 Among the GMLv2-based approaches, prediction-level aggregation with “Mono Pred-level+GMLv2” achieves an overall PCC/SCC of 0.7464/0.52490.7464/0.5249, while “Group Pred-level+GMLv2” improves these to 0.8131/0.77310.8131/0.7731. This improvement indicates that preserving spatially related channel structure is substantially more effective than treating each channel independently. Both approaches, however, require independent backbone inference for each channel or spatial group, with predictions combined only after perceptual processing. At the signal level, ITU Downmix achieves 0.8623/0.78390.8623/0.7839, substantially outperforming Binauralizer at 0.8045/0.75640.8045/0.7564. This suggests that the representation produced by ITU downmixing is more compatible with the pretrained GMLv2 representation than that produced by binauralization. However, fixed signal-domain transformations cannot adapt the contribution of different spatial groups to the input or its degradation characteristics. Among the learned approaches, “Latent Agg+GMLv2” further improves the overall correlation to 0.8875/0.79950.8875/0.7995. This improvement demonstrates the benefit of learning spatial-group contributions rather than relying on a fixed signal-domain transformation. However, latent aggregation still requires a separate GMLv2 inference for each spatial group, combining the groups only after perceptual processing. The feature-level FGAtt achieves the strongest overall performance, with PCC/SCC of 0.9063/0.83030.9063/0.8303. It improves over latent aggregation in both PCC and SCC and consistently outperforms it across all five evaluation sets. Unlike latent aggregation, FGAtt combines the spatial-group representations before perceptual processing, allowing the frozen GMLv2 backbone to jointly process the fused representation with a single backbone inference. Fig. 1 visualizes the overall correspondence between predicted and ground-truth MUSHRA scores for the GMLv2-based approaches. The closer alignment with the identity function is consistent with the quantitative results, with FGAtt showing the strongest overall agreement across the evaluated samples. The per-set results further show the complementary behavior of ITU Downmix and FGAtt. ITU Downmix approach performs particularly strongly on T1–T3, with PCC/SCC of 0.9247/0.87590.9247/0.8759, 0.9213/0.89160.9213/0.8916, and 0.9488/0.88930.9488/0.8893, respectively. FGAtt achieves 0.9001/0.85410.9001/0.8541, 0.9207/0.88800.9207/0.8880, and 0.9106/0.88090.9106/0.8809 on the same sets. On T4 and T5, however, ITU Downmix drops to 0.8493/0.78440.8493/0.7844 and 0.7976/0.68910.7976/0.6891, whereas FGAtt achieves 0.9142/0.78890.9142/0.7889 and 0.9415/0.86160.9415/0.8616. Thus, while ITU Downmix performs strongly on T1–T3, FGAtt provides more consistent performance across the five evaluation sets and achieves higher correlations on T4 and T5. The stronger performance of FGAtt on T4 and T5 is consistent with the benefit of adaptive feature-level integration when fixed signal-domain transformations may not adequately preserve the relative contributions of different spatial groups. ITU downmixing combines multiple channels before perceptual analysis, potentially reducing spatially localized information. In contrast, FGAtt adapts the contribution of each spatial group independently at each Gammatone feature band, allowing the representation presented to GMLv2 to retain different spatial contributions across the perceptual feature space. While the attention weights do not directly identify individual distortion locations, this adaptive representation provides a mechanism for the frozen backbone to jointly exploit complementary information from multiple spatial groups. Overall, the results demonstrate that the level at which multichannel information is integrated substantially affects the effectiveness of extending pretrained GMLv2. Prediction-level and latent-level approaches preserve the backbone’s perceptual processing but require independent inference for each spatial group, while signal-level approaches avoid this cost at the expense of relying on a fixed transformation. FGAtt combines the advantages of learned integration and joint processing by adaptively fusing spatial groups before perceptual processing, while keeping the GMLv2 backbone frozen. Figure 1: Overall correlation between predicted and ground-truth MUSHRA scores across GMLv2 spatial aggregation strategies. The dashed line denotes the identity function. Ablation of Attention Granularity: Having established the effectiveness of feature-level adaptation, we next examine the granularity of attention in FGAtt. As described in Sec. 2, FGAtt estimates a separate attention weight for each spatial group at each Gammatone feature band, while applying the same weight to the four GML-compatible components within the group. We first remove the feature-band dimension, assigning a single weight to each spatial group across all bands. This reduces PCC/SCC from 0.9063/0.83030.9063/0.8303 to 0.8912/0.81280.8912/0.8128, indicating that allowing spatial-group contributions to vary across feature bands is beneficial. We then introduce component-level attention within each spatial group, allowing the four GML-compatible components to receive separate weights in addition to the band-specific group weights. This configuration achieves a PCC/SCC of 0.8803/0.80450.8803/0.8045, lower than FGAtt. The results suggest that feature-band-dependent spatial-group weighting captures the key benefit of the proposed feature-level adaptation, while further decomposing the GML-compatible representation within each group does not provide additional benefit in this setting. Overall, the ablation supports preserving the internal GML-compatible structure while learning feature-band-dependent spatial-group attention. 4 Conclusion We studied how a pretrained perceptual model can be extended to 5.1-channel audio by comparing signal-, prediction-, latent-, and feature-level integration. Beyond fixed signal- and prediction-level strategies, we proposed two learned methods: latent-level aggregation, which weights representations from independently processed spatial groups, and Feature-Band Group Attention (FGAtt), which adaptively fuses spatial-group features before perceptual processing. Across five 5.1-channel test sets, FGAtt achieved the strongest overall performance using a single frozen-backbone inference. These results show that the integration stage is critical and that feature-level adaptation enables effective reuse of pretrained perceptual knowledge. The framework can extend to richer formats through appropriate channel grouping and backbone-compatible representations. References [1] J. Alakuijala, M. Bruse, S. Boukortt, J. M. Coldenhoff, and M. Cernak (2025) Zimtohrli: an efficient psychoacoustic audio similarity metric. arXiv preprint arXiv:2509.26133. Cited by: §3.2, §3.2, Table 1. [2] D. Barry, D. S. Panah, A. Ragano, J. Skoglund, and A. Hines (2025) Binamix–a Python library for generating binaural audio datasets. arXiv preprint arXiv:2505.01369. Cited by: §3.1, Table 1, Table 1. [3] A. Biswas and G. Jiang (2022) Stereo InSE-NET: stereo audio quality predictor transfer learned from mono InSE-NET. In 153rd AES Convention, Cited by: §1. [4] I. Choi, B. G. Shinn-Cunningham, S. B. Chon, and K. Sung (2008) Objective measurement of perceived auditory quality in multichannel audio compression coding systems. Journal of the Audio Engineering Society 56 (1/2), p. 3–17. Cited by: §1, §2.2. [5] P. Delgado and J. Herre (2023) Design choices in a binaural perceptual model for improved objective spatial audio quality assessment. In 155th AES Convention, Cited by: §1. [6] J. Herre, K. Kjörling, J. Breebaart, C. Faller, S. Disch, H. Purnhagen, J. Koppens, J. Hilpert, J. Rödén, W. Oomen, K. Linzmeier, and K. S. Chong (2008) MPEG Surround—the ISO/MPEG standard for efficient and compatible multichannel audio coding. Journal of the Audio Engineering Society 56 (11), p. 932–955. Cited by: §3.1. [7] A. Hines, F. O’Gorman, F. Lim, J. Skoglund, M. Chinen, and N. Gureev (2020) ViSQOL v3: an open source production ready objective speech and audio metric. In 2020 Twelfth International Conference on Quality of Multimedia Experience (QoMEX), Cited by: §1, Table 1, Table 1. [8] A. Hines, J. Skoglund, A. C. Kokaram, and N. Harte (2015) ViSQOL: an objective speech quality model. EURASIP Journal on Audio, Speech, and Music Processing 2015 (1), p. 13. Cited by: §1, §3.2. [9] ITU-R (1994) Recommendation BS.775: multichannel stereophonic sound system with and without accompanying picture. Technical report Technical Report ITU-R BS.775, International Telecommunication Union. Cited by: §2.2, §3.1, Table 1, Table 1. [10] ITU-R (2001) Method for objective measurements of perceived audio quality. Technical report Technical Report BS.1387, International Telecommunication Union. Cited by: §1. [11] ITU-R (2001) Method for the subjective assessment of intermediate quality level of audio systems. Technical report Technical Report BS.1534, International Telecommunication Union. Cited by: §1. [12] G. Jiang, L. Villemoes, and A. Biswas (2023) Generative Machine Listener. In 155th AES Convention, Cited by: §1. [13] S. Kämpf, J. Liebetrau, S. Schneider, and T. Sporer (2010) Standardization of PEAQ-MC: extension of ITU-R BS.1387-1 to multichannel audio. In Audio Engineering Society Conference: 40th International Conference: Spatial Audio, Cited by: §1. [14] I. Loshchilov and F. Hutter (2019) Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), External Links: Link Cited by: §3.1. [15] PQEvalAudio-v1r0 - TSP lab software. Note: https://w.mmsp.ece.mcgill.ca/Documents/Software/(Date last accessed September 2026) Cited by: Table 1, Table 1. [16] V. Raj, K. Gouthaman, S. Gehlot, L. Villemoes, and A. Biswas (2026) Enhanced Generative Machine Listener. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 15147–15151. Cited by: §1, §1, §1, §2.1, §2.2, §3.1. [17] J. Seo, S. B. Chon, K. Sung, and I. Choi (2013) Perceptual objective quality evaluation method for high quality multichannel audio codecs. Journal of the Audio Engineering Society 61 (7/8), p. 535–545. Cited by: §1. [18] T. Thiede, W. C. Treurniet, R. Bitto, C. Schmidmer, T. Sporer, J. G. Beerends, C. Colomes, M. Keyhl, G. Stoll, K. Brandenburg, and B. Feiten (2000) PEAQ—the ITU standard for objective measurement of perceived audio quality. Journal of the Audio Engineering Society 48 (1/2), p. 3–29. Cited by: §1, §3.2, §3.2.