Detection of hypernasality based on vowel space area

Abstract

This study proposes a method for differentiating hypernasal-speech from normal speech using the vowel space area (VSA). Hypernasality introduces extra formant and anti-formant pairs in vowel spectrum, which results in shifting of formants. This shifting affects the size of the VSA. The results show that VSA is reduced in hypernasal-speech compared to normal speech. The VSA feature plus Mel-frequency cepstral coefficient feature for support vector machine based hypernasality detection leads to an accuracy of 86.89% for sustained vowels and 89.47%, 90.57%, and 91.70% for vowels in contexts of high pressure consonants /k/, /p/, and /t/, respectively.

Which Utterance Types Are Most Suitable to Detect Hypernasality Automatically?

“…Based on previous evidence, the following selection of features were computed from each utterance: MFCC coefficients, the first three formants together with their bandwidths (BW) and distances and the VLTHTR [6,10,22,23]. The 13-dimensional MFCC features were calculated using moving Hamming windows.…”

“…In the past there have been many proposals to evaluate HN based on acoustic information. While the results in terms of accuracy are generally excellent, most studies have used only a limited number of utterance types, such as sustained vowels [6][7][8][9][10][11]. This is not clearly compatible with standard clinical recommendations [2], which strongly recommend that patients are evaluated using a variety of phonemes and utterances with varying complexity (e.g., CAPS-A protocol, [2,12]).…”

“…Classification algorithms include Random Forest (RF), Support Vector Machine (SVM) or Artificial Neural Network (ANN). Acoustic features used in previous studies include, among others, Mel Frequency Cepstrum Coefficients (MFCCs), the Voice Low Tone to High Tone Ratio (VLTHTR) and the vowel formants and their bandwidth [6][7][8][9][10][11]. In most studies using this approach the fragments analyzed were either sustained vowels [6,8] or vowel fragments that had been annotated manually in words or sentences [7,[9][10][11].…”

“…Acoustic features used in previous studies include, among others, Mel Frequency Cepstrum Coefficients (MFCCs), the Voice Low Tone to High Tone Ratio (VLTHTR) and the vowel formants and their bandwidth [6][7][8][9][10][11]. In most studies using this approach the fragments analyzed were either sustained vowels [6,8] or vowel fragments that had been annotated manually in words or sentences [7,[9][10][11]. Only a few studies have used complex utterances to automatically evaluate nasality [17][18][19].…”

Hypernasality Severity Detection Using Constant Q Cepstral Coefficients

“…The important works on hypernasality detection are based on Teager energy operator (TEO) based feature [11], TEO feature with frequency cepstral coefficient (MFCC) feature [12], pitch-adaptive MFCC feature [13], linear prediction cepstral coefficient (LPCC) feature [14] and features extracted from high spectral resolution group delay spectrum [4] and zero time windowing technique [15], [16]. Besides that the feature set obtained from acoustic, noise and cepstral analysis, nonlinear dynamic and entropy measurements [17], [18], [19], based on energy distribution [20], [21] and using vowel space area (VSA) [22] are also used. The hypernasality detection is also done using recorded sentences speech database using jitter, shimmer, MFCC, bionic wavelet transform entropy and bionic wavelet transform energy features.…”