Nasal Air Emission in Sibilant Fricatives of Cleft Lip and Palate Speech

Sishir Kalita, Protima Nomo Sudro, S.r. Mahadeva Prasanna, S. DandapatInterspeech 2019

TL;DR

This paper investigates the acoustic characteristics of sibilant fricatives in individuals with cleft lip and palate (CLP) who exhibit nasal air emission (NAE). They propose a Support Vector Machine (SVM) based classification system to distinguish NAE-distorted fricatives from normal ones, achieving high accuracy by analyzing spectral moments and temporal variations in auditory excitation.

What Problem Does This Paper Address?

Cleft lip and palate (CLP) is a congenital condition that can lead to velopharyngeal dysfunction (VPD), a common cause of nasal air emission (NAE) in speech. NAE occurs when air escapes through the nasal cavity during the production of oral sounds, particularly fricatives, resulting in a distorted and unpleasant speech quality that significantly impacts intelligibility. Current methods for assessing NAE rely heavily on subjective perceptual evaluations by speech-language pathologists (SLPs), which can be inconsistent and depend on the clinician's expertise. While instrumental methods exist, they often require more patient involvement. This paper aims to bridge this gap by developing an objective, technology-based approach to identify and classify NAE-distorted fricatives using acoustic analysis, thereby assisting SLPs in diagnosis and therapy.

Methods

The study focuses on analyzing voiceless sibilant fricatives (/s/ and /f/) from individuals with CLP exhibiting NAE and a control group with normal speech. The core methodology involves two main approaches:

  1. Static Spectral Analysis: This involves analyzing the spectral distribution of the fricatives using four spectral moments: mean (M1), standard deviation (M2), skewness (M3), and kurtosis (M4). These moments characterize the shape and spread of the spectral energy.
  2. Dynamic Spectral Analysis: To capture temporal variations, the researchers employ a psychoacoustic measure called the ERBN-number (Equivalent Rectangular Bandwidth number), which represents the dominant frequency of the auditory excitation over time. This is computed using a gammatone filterbank. Additionally, they extract auditory excitation features (AEF) using the Discrete Cosine Transform (DCT) on the spectral envelope, along with their temporal derivatives (ΔAEF and ΔΔAEF), resulting in a 39-dimensional feature vector.

Finally, a Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF) kernel is trained and tested using these extracted features to differentiate between normal and NAE-distorted fricatives.

Datasets

The speech data was collected from:

  • 20 children with repaired cleft lip and palate (CLP) exhibiting NAE.
  • 20 age and gender-matched children with normal speech and language characteristics as controls.

All participants were native Kannada speakers, aged 7-12 years. The data was recorded in a soundproof room using a directional microphone at a sampling frequency of 44 kHz. The fricative portions (/s/ and /f/) were manually extracted from CVCV utterances.

Key Insights and Ideas

  • Spectral Shift in NAE: NAE-distorted fricatives show a significant concentration of spectral energy in lower frequency ranges compared to normal fricatives. For /s/, this shift is from above 4 kHz to between 2-5 kHz, and for /f/, from 2-8 kHz to 1.5-8 kHz. This suggests that the nasal cavity acts as an additional resonator, altering the sound's spectral profile.
  • Increased Temporal Variability: The temporal trajectories of the ERBN-number exhibit greater deviations and less stability in NAE-distorted fricatives compared to normal ones. This indicates less consistent articulatory control or airflow dynamics during speech production.
  • Skewness as a Discriminator: Spectral skewness (M3) proves to be a significant indicator, with NAE-distorted fricatives showing a more positive spectral slope, implying lower spectral density in higher frequencies.
  • SVM Efficacy: The SVM classifier demonstrates strong performance in distinguishing between normal and NAE-distorted fricatives, suggesting that acoustic features can effectively capture the subtle differences caused by NAE.

Key Results

  • Spectral Moments:
    • For /s/, spectral mean (M1) showed a significant difference (p<0.001) between normal and NAE-distorted speech. Skewness (M3) was also a significant differentiator.
  • For /f/, spectral mean (M1) was highly significant (p<0.001), and standard deviation (M2) also showed a significant difference (p<0.001).
  • Classification Performance:
    • Frame-level: The SVM achieved an overall accuracy of 89.95% for /s/ and 94.85% for /f/.
  • Utterance-level: Accuracies improved to 93.27% for /s/ and 97.81% for /f/.
  • The higher accuracy for /f/ might be attributed to more severe NAE distortion in the collected /f/ samples.

Limitations and Future Work

  • Database Size: The study acknowledges that the current work is preliminary and based on a relatively small database.
  • Future Directions:
    • Expand the study to include a larger database.
  • Investigate other pressure consonants (stops and affricates) affected by NAE.
  • Analyze different severity levels of NAE (mild vs. severe) to gain deeper insights into acoustic characteristics.
  • Explore the acoustic characteristics of other speech sounds distorted by NAE.

Summary updated

BibTeX
@inproceedings{kalita2019nasal,
  author = {Sishir Kalita and Protima Nomo Sudro and S.R. Mahadeva Prasanna and S. Dandapat},
  title = {{Nasal Air Emission in Sibilant Fricatives of Cleft Lip and Palate Speech}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure