TL;DR
This paper introduces a novel, data-driven system for automatically assessing voice and speech intelligibility in patients who have undergone laryngeal cancer treatment. By modeling acoustic features of read speech, the system achieves high correlations with expert human evaluations, offering a promising objective support tool for clinical assessment and therapy monitoring.
What Problem Does This Paper Address?
Laryngeal cancer significantly impacts a person's voice and communication abilities, affecting their quality of life and economic prospects. While clinical evaluation of speech and voice quality after treatment often relies on subjective assessments by expert raters, this method is time-consuming, prone to individual bias, and can suffer from inter-rater variability. Existing automatic acoustic analysis methods often struggle with the highly pathological and aperiodic voices of these patients, particularly when analyzing fundamental frequency (F0) from sustained vowels. This paper addresses the need for a robust, objective, and automatic system that can reliably assess voice and speech quality, specifically intelligibility, in post-laryngeal cancer patients using connected speech.
Methods
The core of the proposed system is a statistical speaker modeling approach. Instead of relying on traditional F0 analysis, it utilizes Mel-frequency Cepstral Coefficients (MFCCs) extracted from short segments of read speech. These MFCCs are used to build speaker-specific Gaussian Mixture Models (GMMs), which represent the acoustic characteristics of an individual's voice.
Here's a breakdown of the key steps:
- Feature Extraction: MFCCs are computed from the speech signal, capturing spectral information. These features are then transformed into a higher-dimensional representation (36 dimensions) by including first and second-order derivatives.
- Speaker Modeling: A Universal Background Model (UBM) is trained on data from healthy speakers. This UBM serves as a reference. For each pathological speaker, their GMM is adapted from the UBM using Maximum A Posteriori (MAP) adaptation. This process creates a speaker-specific GMM.
- Supervector Representation: To reduce dimensionality and create a computationally efficient representation, the mean vectors of the Gaussian densities within each speaker's GMM are concatenated to form a "GMM-based supervector."
- Regression for Evaluation: Support Vector Regression (SVR) is employed to model the relationship between the GMM-based supervectors and the perceptual evaluation scores (strain, voice quality, intelligibility) provided by human experts. Separate SVR models are trained for each evaluation criterion.
This approach allows for the evaluation of connected speech, overcoming limitations of methods that rely on sustained vowels.
Datasets
The study utilized two distinct datasets of speech samples from patients who had undergone laryngeal cancer treatment:
- Tracheoesophageal (TE) Speakers: 54 patients who had undergone total laryngectomy and were equipped with a Provox shunt valve. These speakers were recorded on average 2.4 years after surgery.
- Partial Laryngectomy (PL) Speakers: 77 patients who had undergone partial removal of the larynx, preserving at least one vocal fold. These speakers were recorded at least 1 year after surgery.
All speakers read a standardized German text, "The North Wind and the Sun." The recordings were made using a headset with a 16 kHz sampling frequency.
Key Insights and Ideas
- Moving Beyond F0: The system's success in analyzing highly pathological voices, which often have unreliable F0 extraction, highlights the effectiveness of using spectral features (MFCCs) and GMM-based modeling as an alternative.
- Connected Speech Advantage: Evaluating connected speech, rather than just sustained vowels, provides a more ecologically valid assessment of communication abilities, which is crucial for intelligibility.
- Data-Driven Robustness: The purely data-driven approach, without relying on explicit linguistic or phonetic knowledge, makes the system adaptable and potentially robust to variations in speech patterns.
- Objective Support for Subjective Evaluation: The high correlation between the automatic system's scores and expert ratings suggests it can serve as a valuable objective tool to complement subjective clinical assessments, reducing bias and saving time.
Key Results
The study reported strong correlations between the automatic system's evaluations and the mean perceptual scores from expert listeners:
- TE Speakers: Correlations ranged from $r = 0.81$ for strain to $r = 0.83$ for intelligibility.
- PL Speakers: Correlations ranged from $r = 0.80$ for strain to $r = 0.84$ for voice quality.
These correlations were found to be in the same range as the inter-rater reliability among the human experts themselves, indicating the system's performance is comparable to human judgment. The authors also noted that the number of Gaussian densities used in the GMMs (tested up to 256) did not significantly impact the results, suggesting a degree of flexibility in model complexity.
Limitations and Future Work
- Limitations:
- The study acknowledges that perceptual evaluations, even among experts from the same department, showed some divergence, which is a known challenge in subjective assessments.
- The current system's acoustic information is extracted without explicit temporal order, which the authors suggest might be a limitation for certain aspects of speech analysis.
- Future Work:
- The authors suggest future work could explore incorporating temporal information into the acoustic modeling.
- They also envision applications such as interactive training tools for patients and further exploration of the system's utility for other voice and speech disorders.
- The system's robustness to reading errors and its potential for more detailed phonetic analysis are highlighted as areas for future investigation.
Summary updated