TL;DR
This paper introduces an adaptive, time-varying, bi-directional linear predictive model to analyze disordered speech. By capturing long-term redundancies in speech signals, this method provides a more robust measure of vocal dysperiodicity, correlating well with perceived hoarseness, even in continuous speech.
What Problem Does This Paper Address?
Assessing disordered speech, particularly hoarseness, often relies on acoustic markers. The signal-to-noise ratio (SNR) is a common metric, but traditional methods using short-term linear prediction can be influenced by speaker-specific characteristics unrelated to vocal disorders. While long-term prediction (LTP) improves this by modeling cycle-to-cycle redundancies, it struggles with non-stationary speech and transitions between phonetic segments. Bi-directional LTP (BLTP) addresses some of these issues by considering both forward and backward predictions, but it still assumes local stationarity. This paper tackles the challenge of analyzing disordered speech in a more general, time-varying context, aiming to capture vocal dysperiodicity without assuming local stationarity, thus improving the accuracy and robustness of SNR estimation for continuous speech.
Methods
The core innovation is an adaptive, time-varying, bi-directional linear predictive model. Unlike previous methods that assumed local stationarity, this model continuously updates its prediction coefficients for each speech sample.
- Time-Varying Long-Term Prediction: The model predicts a current speech sample
x(n)based on past samplesx(n-P(n)-i), whereP(n)is a time-dependent prediction distance. This allows for capturing the varying cycle length in speech. - Bi-Directional Prediction: To overcome issues at the beginning of recordings and segment boundaries, the model performs both forward and backward predictions. The final prediction error is chosen to be the minimum of the two. This effectively avoids predicting across phonetic segment boundaries where cycle shapes naturally change.
- Adaptive Estimation: The model coefficients are estimated adaptively using a recursive least squares (RLS) algorithm. This allows for real-time updates and handling of non-stationary signals.
- SNR Calculation: The paper defines an acoustic marker, the signal-to-noise ratio (SNR), calculated using the original signal and the minimal prediction error obtained from the bi-directional LTP. This SNR is proposed as a measure of vocal dysperiodicity.
Datasets
The paper evaluates its method on two corpora:
- Sustained Vowel Corpus: 1-second stationary fragments of the vowel [a] from 38 healthy and 51 dysphonic speakers. Hoarseness was judged visually from spectrograms by five judges, resulting in an overall perceived hoarseness score between 0 and 20.
- Continuous Speech Corpus: A subset of the French sentence "il est sorti avant le jour" uttered by a female speaker. These signals were labeled with increasing degrees of hoarseness: "modal," "rough 1," "rough 2," "rough 3," and "whisper."
Key Insights and Ideas
- Capturing Vocal Dysperiodicity: The core idea is that vocal dysperiodicity (irregularities in voice production) manifests as increased prediction error in long-term predictive models. The more irregular the voice, the harder it is to predict future samples from past ones.
- Robustness to Non-Stationarity: By making the model adaptive and time-varying, the authors overcome a major limitation of previous LTP methods, allowing for accurate analysis of continuous speech with natural variations.
- Bi-Directionality for Boundary Handling: The bi-directional approach is crucial for avoiding spurious prediction errors at the start of a signal or when transitioning between different sounds (phonemes). It ensures that the prediction error is primarily due to vocal pathology, not linguistic changes.
- SNR as a Reliable Metric: The study demonstrates a strong correlation between the proposed SNR metric and the perceived degree of hoarseness, suggesting its utility as an objective measure for clinical assessment.
Key Results
- Correlation with Hoarseness: For the sustained vowel corpus, the calculated SNR values showed a strong negative correlation (Spearman's rank correlation coefficient of -0.8) with the perceived degree of hoarseness. As hoarseness increased, SNR decreased.
- Continuous Speech Performance: The method successfully analyzed continuous speech, with SNR values aligning with subjective quality labels. The "modal" (normal) speech had a high SNR (25.3 dB), while "whisper" (highly dysphonic) had a low SNR (10.4 dB).
- Visual Evidence: Figure 1 shows a speech signal and its corresponding prediction error, visually illustrating how dysperiodicity leads to larger prediction errors.
Limitations and Future Work
- Prediction Distance Optimization: The paper notes that the prediction distance
P(n)was determined via cross-correlation. They suggest that results might be improved by using an interpolation procedure to find a more accurate prediction distance. - Real-time Tracking: The authors propose that a potential future application is the real-time tracking of evolving vocal dysperiodicity marker values in connected speech, leveraging the adaptive nature of their model.
Summary updated