Introduction
In speech technology, accurately assessing the quality of an audio signal as perceived by a human listener is critical. While intrusive quality assessment (QA) compares a distorted signal to a clean reference, non-intrusive QA must evaluate the signal in isolation. This is a significantly more challenging task, often plagued by overfitting and a lack of transparency regarding the model's confidence in its predictions.
Researchers have addressed these issues by developing a system that combines Modulation Spectrum features with Bayesian Multivariate Adaptive Regression Splines (BMARS). This approach not only achieves high accuracy but also provides "credible intervals" that indicate how much the system trusts its own quality scores.
Methodology
Feature Extraction: The Modulation Spectrum
To mimic human auditory perception, the system extracts the temporal envelope of speech across various frequency bands. It specifically looks at:
- Articulation Band Powers (ABP): Energy associated with speech production.
- Non-articulation Band Powers (NBP): Energy typically introduced by noise or distortion.
By calculating the global mean and variance of these features over an entire utterance, the model captures a statistical fingerprint of the audio quality.
Modeling: Bayesian MARS (BMARS)
While standard MARS uses flexible mathematical curves (splines) to fit data, the Bayesian implementation treats parameters as probability distributions. This provides two major advantages: it naturally penalizes overly complex models to prevent overfitting, and it allows for the quantification of uncertainty.
Quantifying Uncertainty
A unique strength of this Bayesian approach is the generation of credible intervals. Most models provide a single Mean Opinion Score (MOS) without context. BMARS, however, flags when it is "unsure"—for instance, when encountering a type of distortion not present in its training data.
As shown in the experimental results, increasing the training data four-fold can reduce these intervals by approximately 66%. This "confidence meter" allows the system to identify unfamiliar distortions, such as specific vehicle noises, by widening the interval and signaling that the prediction may be less reliable.
Performance and Results
The model was validated using the ITU-T P.Sup23 database, which contains over 1,300 utterances across multiple languages and laboratory conditions. The BMARS model consistently outperformed existing standards.
The proposed system achieved an average Pearson correlation of 0.892, surpassing the then-standard ITU-T P.563 algorithm (0.876) and non-Bayesian MARS (0.873). This demonstrates that the Bayesian framework is not only more informative due to its uncertainty metrics but also more accurate in its primary task of quality prediction.
Conclusion and Future Work
By integrating auditory-inspired features with a robust Bayesian regression framework, this research provides a more reliable tool for automated speech quality assessment. While the current model is optimized for narrow-band speech (8 kHz), future iterations could expand to wide-band audio. Additionally, while the training process is computationally intensive due to Markov Chain Monte Carlo sampling, the resulting model is highly efficient for real-time predictions, making it a viable candidate for monitoring modern telecommunications networks.
Summary updated