Introduction
Assessing speech intelligibility is a critical task for clinicians monitoring disease progression or evaluating therapy effectiveness in patients with speech pathologies, such as Cerebral Palsy. While subjective listening tests remain the clinical "gold standard," they are resource-intensive, time-consuming, and susceptible to listener bias.
Existing automatic assessment methods often face significant hurdles. Data-driven models like Automatic Speech Recognition (ASR) or iVector systems require massive datasets of healthy speech for training. Alternatively, state-of-the-art measures like P-ESTOI rely on Dynamic Time Warping (DTW) to align patient speech with a healthy reference word-for-word. This alignment is computationally expensive and fails in "phonetically unbalanced" scenarios where the patient and the reference speaker are not saying the same words.
Methodology: Subspace-Based Analysis
This research proposes a computationally efficient, non-blind approach based on spectral subspace analysis. The core hypothesis is that intelligible speech occupies a specific mathematical "space" of spectral patterns, and pathological speech can be characterized by its geometric drift away from this healthy subspace.
The process begins with feature extraction using one-third octave band analysis to mimic human auditory processing. The healthy spectral bases ($U_H$) are computed using two linear decomposition methods: Principal Component Analysis (PCA) and Approximate Joint Diagonalization (AJD). While PCA averages correlations across speakers, AJD finds a common basis that represents multiple healthy speakers simultaneously, offering greater robustness to diverse vocal characteristics.
An "L-curve" method automatically selects the optimal number of bases ($B$) to balance reconstruction accuracy against noise. Finally, the patient's subspace is compared to the healthy subspace by treating them as points on a Grassmann manifold and calculating the Chordal distance. This distance serves as a direct proxy for the intelligibility score.
Experimental Setup and Results
The method was validated using the Universal Access (UA) speech database, featuring 15 patients with Cerebral Palsy and 13 healthy control speakers. The dataset included 763 isolated words with subjective human intelligibility ratings ranging from 2% to 95%.
The results demonstrate that the subspace method is highly effective and robust:
- High Correlation: The proposed measure achieved a Pearson correlation ($R$) of up to 0.83 with subjective human ratings, approaching the performance of more complex alignment-dependent models.
- Computational Efficiency: By eliminating the need for DTW, the method is significantly faster than P-ESTOI, making it suitable for real-time clinical applications.
- Phonetic Independence: Most notably, the AJD-based method maintained a high correlation ($R = 0.78$) even when the patient and the healthy reference spoke entirely different words. This suggests that the underlying "spectral signature" of intelligibility is consistent enough to be measured without strict phonetic alignment.
Conclusion
Spectral subspace analysis offers a promising path toward fast, data-efficient, and phonetically independent automatic speech assessment. While the current model focuses on spectral patterns and does not yet account for temporal distortions (rhythm and timing), future work incorporating temporal modulation features could further enhance accuracy. This approach provides a viable tool for clinicians to obtain objective intelligibility scores without the overhead of traditional listening tests or complex alignment algorithms.
Summary updated