Meta Audiobox Aesthetics Framework

• Meta Audiobox Aesthetics is a multidimensional framework that assesses synthetic audio by decomposing quality into production, complexity, enjoyment, and usefulness.
• It leverages a Transformer-based neural model with fine-grained axis prediction to enhance evaluation and guide reinforcement learning in text-to-audio and music generation.
• The framework supports robust benchmarking across diverse audio modalities, ensuring improved quality assessment while noting challenges in over-optimization and diversity loss.

Meta Audiobox Aesthetics, as formalized in recent research, encompasses a comprehensive, multidimensional framework for evaluating, analyzing, and generating audio artifacts—spanning speech, music, and sound effects—with a focus on both objective production values and subjective listener response. This approach integrates aesthetic axes tailored to human perceptual experience and provides neural models and annotation protocols for automatic, fine-grained quality assessment. The framework has influenced the development of benchmarks, generative models, and reinforcement learning pipelines for advancing the aesthetic quality of synthetic audio content across modalities.

1. Theoretical Foundations and Aesthetic Axes

Meta Audiobox Aesthetics departs from traditional, uni-dimensional evaluation approaches (such as overall Mean Opinion Score or FAD) by proposing a four-axis decomposition of human auditory aesthetic judgment (Tjandra et al., 7 Feb 2025). These axes are:

• Production Quality (PQ): Technical fidelity, clarity, dynamics, frequency balance, spatialization.
• Production Complexity (PC): Scene density or number of concurrent audio components.
• Content Enjoyment (CE): Emotional impact, artistic expression, and overall listener pleasure.
• Content Usefulness (CU): Potential for the audio to serve as usable source material for creative tasks.

This orthogonalization allows the annotation protocol to elicit structured, interpretable feedback from listeners, thereby avoiding corpus effects and ambiguities associated with ambiguous aggregate labels.

2. Model Architecture and Prediction Methodology

The automatic assessment system developed for Meta Audiobox Aesthetics is a no-reference, utterance-level predictor based on a Transformer architecture using a WavLM encoder (12 Transformer layers; 768d hidden) (Tjandra et al., 7 Feb 2025).

Embedding Extraction and Prediction

Let $h_{l,t}$ denote the t-th hidden state of the l-th WavLM layer, with per-layer learnable scalar weight $w_l$:

$z_l = w_l$

$$\hat{e} = \sum_{l=1}^L \sum_{t=1}^T ( h_{l,t} \cdot z_l )$$

$e = \frac{\hat{e}}{||\hat{e}||_2}$

The normalized embedding $e$ is input to MLP predictors (linear layers, LayerNorm, GeLU) to output axis-specific scores for PQ, PC, CE, CU. The loss comprises mean squared error (MSE) and mean absolute error (MAE):

$$\mathcal{L} = \sum_{a \in \{\mathrm{PQ}, \mathrm{PC}, \mathrm{CE}, \mathrm{CU}\}} \left( (y_a - \hat{y}_a)^2 + |y_a - \hat{y}_a| \right)$$

Data and Preprocessing

Input waveforms are resampled to 16 kHz; 10 s audio chunks are sampled during training, with sliding window inference and weighted averaging for final predictions.

3. Evaluation Protocols and Challenge Results

The AudioMOS Challenge 2025 (Huang et al., 1 Sep 2025) institutionalized the Meta Audiobox Aesthetics axes as the foundation of Track 2, evaluating algorithmic predictors on synthetic samples from text-to-speech, text-to-audio, and text-to-music systems. Metrics included both sample-level and system-level:

Metric	Formula (brief)
Mean Squared Error (MSE)	$(1/N) \sum ( \hat{y}_i - y_i )^2$
Linear Correlation (LCC)	$$\frac{\sum (x_i-\bar{x})(y_i-\bar{y})}{\sqrt{\sum (x_i-\bar{x})^2 \sum (y_i-\bar{y})^2}}$$
Spearman Rank Corr (SRCC)	$1 - \frac{6 \sum d_i^2}{N(N^2-1)}$
Kendall's Tau	Rank-based ordinal coefficient

System-level SRCC was the primary ranking metric. Top submissions employed WavLM, CLAP, ensemble pipelines, feature fusion, and custom regression/rank-consistent losses, demonstrating that fine-grained axis prediction outperformed traditional MOS-based schemes. Data scaling beyond the curated AES-Natural dataset yielded diminishing returns, highlighting the importance of annotation quality and diversity.

4. Practical Implications in Generation and Reinforcement Learning

The practical utility of Meta Audiobox Aesthetics extends to text-to-audio, text-to-music, and style transfer systems:

• Direct integration of per-axis scores as feedback signals in RL: The SMART system (Jonason et al., 23 Apr 2025) demonstrated that reinforcement learning fine-tuning of a symbolic music generator with Content Enjoyment ratings from the Meta Audiobox Aesthetics model improved subjective ratings and altered low-level features (note count, polyphony, pitch range, scale consistency).
• Reward over-optimization reduces output diversity: Omission of KL penalty in RL led to reward hacking, where diversity collapsed and outputs became highly repetitive despite higher prediction scores, paralleling findings in RL for LLMing (Jonason et al., 23 Apr 2025).

5. Relationship to Historical and Theoretical Frameworks

Meta Audiobox Aesthetics' decomposition aligns with historical treatments of auditory display aesthetics in the fields of sonification and program auralization (Vickers et al., 2013, Vickers et al., 2013, Vickers, 2013). Early work formalized conceptual dimensions (e.g., indexicality, abstraction), ecological validity, and the convergence of musical and sonificational design. These antecedents recognized that aesthetic judgment is inherently multi-factorial: balancing technical, cognitive, artistic, and utilitarian considerations.

The new framework operationalizes these insights by supplying standardized annotation protocols and neural predictors for the axes most salient to human listeners in synthetic audio evaluation.

6. Impact on Benchmarking and Methodological Advances

Meta Audiobox Aesthetics has enabled robust benchmarking for generative audio modeling, pseudo-labeling, and large-scale data selection. Its predictors facilitate filtering and assessing generative outputs for quality at scale, supporting research in music, TTS, and general sound generation (Tjandra et al., 7 Feb 2025, Huang et al., 1 Sep 2025). The open-source release of both data and pretrained models anchors reproducible, extensible research in the multi-aspect automatic evaluation of audio.

A plausible implication is that adopting multi-axis aesthetics predictors could reduce bias and improve alignment with specific generation objectives (e.g., maximizing Content Usefulness for production libraries or Content Enjoyment for entertainment domains), compared to previous approaches focused solely on global similarity or technical metrics.

7. Limitations and Future Research Directions

While decomposed axis prediction mitigates ambiguity compared to overall MOS and FAD, results demonstrate that axes are not equally predictive for all audio types (e.g., Production Complexity for intelligibility in speech (Tjandra et al., 7 Feb 2025)). Over-optimization for per-axis scores may threaten diversity and longer-term structure, as observed in music RL settings (Jonason et al., 23 Apr 2025). Extensions to longer context, cross-modal (audio–text–image) aesthetics prediction, and adversarial robustness in annotation remain open challenges.

Further development of annotation protocols, fusion with domain-specific metrics, and exploration of axis interdependence are likely to further refine the explanatory and predictive power of Meta Audiobox Aesthetics in universal audio quality assessment and generation.