Learning Voice Source Related Information for Depression Detection

Abstract

During depression neurophysiological changes can occur, which may affect laryngeal control i.e. behaviour of the vocal folds. Characterising these changes in a precise manner from speech signals is a non trivial task, as this typically involves reliable separation of the voice source information from them. In this paper, by exploiting the abilities of CNNs to learn task-relevant information from the input raw signals, we investigate several methods to model voice source related information for depression detection. Specifically, we investigate modelling of low pass filtered speech signals, linear prediction residual signals, homomorphically filtered voice source signals and zero frequency filtered signals to learn voice source related information for depression detection. Our investigations show that subsegmental level modelling of linear prediction residual signals or zero frequency filtered signals leads to systems better than the state-of-the-art low level descriptor based systems and deep learning based systems modelling the vocal tract system information.

Using Speech Production Knowledge for Raw Waveform Modelling Based Styrian Dialect Identification

“…"FC" layer in this architecture contains 100 nodes. The input to the CNNs is a 250ms signal, overlapped by a 10ms shift; these parameters are inspired from earlier works such as [13,17]. The targets to the CNNs are one-hot encodings of the dialects.…”

“…During testing, such posterior probability vectors are averaged across each utterance and a decision is made at the utterance-level based on the highest probability. Our recent work [17] showed that filtering raw speech based on prior knowledge facilitates better task-specific modelling. Along similar lines, apart from directly modelling raw speech and using speed perturbation (SP) as suggested in [11] to partially address data scarcity, we propose to use the following signal processing techniques to extract signals rich in vocal-tract related information: (a) homomorphic filtering and (b) linear prediction based filtering.…”

“…In this direction, we study DID through CNN-based modelling of signals of raw speech that contain vocal-tract related information and through modelling their articulatory differences. Specifically, (a) motivated by the recent works that modelled raw speech directly for specific tasks [13,14,15,16] and their signal-processed variants [17], we study DID using knowledge from source-filter decomposition, and (b) motivated by the acoustic modelling of articulatory differences among phones [18], we study learning such knowledge from raw signals of speech on resource-rich English language and transferring it to the Styrian DID problem.…”

Estimating the Degree of Sleepiness by Integrating Articulatory Feature Knowledge in Raw Waveform Based CNNS

“…We used the raw-speech based CNN framework originally developed for speech recognition [15], and later extended to other tasks such as speaker verification [11], gender identification [8], presentation attack detection [12] or depression detection [5]. In this framework, as illustrated in Figure 1, the network consists of N convolution layers (Conv), maximum pooling (MaxP) and ReLU activations followed by a multilayer perceptron (MLP).…”

“…In the recent years, with the advances in deep learning, approaches have emerged where task-related information are directly learned from raw speech signals using convolutional neural networks (CNNs) in an end-to-end manner, i.e. without any short-term spectral processing [5,11,12,14,18,22,24]. This paper investigates the use of such an approach for the degree of sleepiness estimation.…”