Investigation of self-supervised pre-trained models for classification of voice quality from speech and neck surface accelerometer signals

Prior studies in the automatic classification of voice quality have mainly studied use acoustic speech signal as input. Recently, a few been carried out by jointly using both and neck surface accelerometer (NSA) signals inputs, extracting MFCCs glottal source features. This study examines simultaneously-recorded NSA (breathy, modal, pressed) features derived from three self-supervised pre-trained models (wav2vec2-BASE, wav2vec2-LARGE, HuBERT) SVM well CNNs classifiers. Furthermore, effectiveness is compared feature extraction between waveforms raw for inputs. Using two processing methods (quasi-closed phase (QCP) inverse filtering zero frequency (ZFF)), are estimated signals. The has main goals: (1) to whether improve accuracy conventional (spectrogram, mel-spectrogram, MFCCs, i-vector, x-vector), (2) investigate which modalities (speech vs. NSA) more effective task with model-based features, (3) evaluate deep learning-based CNN classifier can enhance comparison classifier. results revealed that input showed better performance signal. Between accuracies, inputs It was also found HuBERT performed than wav2vec2-BASE wav2vec2-LARGE