A speech quality classifier based on Tree-CNN algorithm that considers network degradations

• Autor(a) principal: Vieira, Samuel Terra
• Data de Publicação: 2020
• Outros Autores: Rosa, Renata Lopes, Zegarra Rodríguez, Demóstenes
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/42433
• topic: Speech quality; Objective metrics; Wireless network; Wired network; Deep learning; Tree Convolutional Neural Network; Voz - Qualidade; Rede sem fio; Rede com fios; Aprendizagem profunda; Redes neurais convolucionais

Many factors can affect the users’ quality of experience (QoE) in speech communication services. The impairment factors appear due to physical phenomena that occur in the transmission channel of wireless and wired networks. The monitoring of users’ QoE is important for service providers. In this context, a non-intrusive speech quality classifier based on the Tree Convolutional Neural Network (Tree-CNN) is proposed. The Tree-CNN is an adaptive network structure composed of hierarchical CNNs models, and its main advantage is to decrease the training time that is very relevant on speech quality assessment methods. In the training phase of the proposed classifier model, impaired speech signals caused by wired and wireless network degradation are used as input. Also, in the network scenario, different modulation schemes and channel degradation intensities, such as packet loss rate, signal-to-noise ratio, and maximum Doppler shift frequencies are implemented. Experimental results demonstrated that the proposed model achieves significant reduction of training time, reaching 25% of reduction in relation to another implementation based on DRBM. The accuracy reached by the Tree-CNN model is almost 95% for each quality class. Performance assessment results show that the proposed classifier based on the Tree-CNN overcomes both thecurrent standardized algorithm described in ITU-T Rec. P.563 and the speech quality assessment method called ViSQOL.