arXiv · 2025 · Preprint
Tjandra et al. (Meta (FAIR)) · → Paper · Demo: ? · Code: ✓
Introduces Audiobox-Aesthetics, a set of no-reference automatic predictors that decompose audio quality into four orthogonal axes (Production Quality, Production Complexity, Content Enjoyment, Content Usefulness) across speech, music, and sound, along with the AES-Natural evaluation dataset and open-source models.
Citation Stub
This paper is not a speech generation paper but is cited by the corpus. See Context in Speech Generation below for why it is relevant.
Context in Speech Generation
Audiobox-Aesthetics provides an alternative to single-dimension MOS-style scoring by decomposing perceived audio quality into four independently trained predictors. Each predictor targets a distinct perceptual axis, enabling more diagnostic quality assessment than a single overall quality score. The model architecture builds on a WavLM-based encoder with learned layer-weighted pooling, making it directly usable as a no-reference, reference-free quality estimator for generated speech and audio. The paper demonstrates the system’s utility for TTS pipeline improvement: using Audiobox-Aesthetics Production Quality scores as text prompt prefixes during training reduces WER and improves subjective pairwise ratings compared to data filtering alone, while retaining the full training corpus.
Wiki Connections
- evaluation-metrics — introduces a four-axis automatic quality scoring framework usable as a no-reference quality evaluator in speech generation pipelines
- subjective-evaluation — the AES-Natural dataset provides human annotations across four perceptual dimensions that extend beyond standard MOS single-axis listening tests
- 2210.13438 (EnCodec) — cited as a neural codec whose output quality is evaluated using the Audiobox-Aesthetics predictors
- 2204.02152 (UTMOS) — cited as a related automatic MOS predictor; Audiobox-Aesthetics extends the single-dimension UTMOS approach to multi-axis assessment
- 2312.15821 (Audiobox) — the speech and audio generation system that Audiobox-Aesthetics was developed alongside and validated on
- 2407.21783 (LLaMA 3) — cited as a backbone component in the broader Audiobox system context