This report covers a robust voice activity detection (VAD) algorithm presented in [1]. The algorithm uses higher order statistics (HOS) metrics of speech signal in linear prediction coding (LPC) residual domain to classify noise and speech frames of a signal. Chapters in this report present voice activity detection problem and analysis of environment issues for VAD, deep HOS based and standard algorithms analysis and a real time HOS based voice activity detector model. New improvements (instantaneous SNR estimation, decision smoothing, adaptive thresholds, artificial neural network) to the proposed algorithm are introduced and performance results of the improved algorithm compared to standard VAD algorithms are presented.

Thesis (M.S.)--Florida State University, 2005. / Advisor: Dr. Anke Meyer-Baese, Florida State University, College of Engineering, Dept. of Electrical Engineering. Title and description from dissertation home page (viewed June 15, 2005). Document formatted into pages; contains vii, 30 pages. Includes bibliographical references.

Submitted by Johnny Rodrigues (johnnyrodrigues@ufcg.edu.br) on 2017-09-19T15:36:01Z No. of bitstreams: 1 Dissertacao_SergioEspinola_CEEI_UFCG.pdf: 59559230 bytes, checksum: 045a4738e365ab656e17da8b2185cb9b (MD5) / Made available in DSpace on 2017-09-19T15:36:01Z (GMT). No. of bitstreams: 1 Dissertacao_SergioEspinola_CEEI_UFCG.pdf: 59559230 bytes, checksum: 045a4738e365ab656e17da8b2185cb9b (MD5) Previous issue date: 2014-03-12 / Estima-se que um terço da força de trabalho humana dependa da voz para realização de seus ofícios. Procedimentos médicos avaliam a qualidade vocal do indivíduo sendo os mais usados aqueles baseados na escuta da voz (subjetivo) ou na inspeção das dobras (ou pregas) vocais por exames sofisticados (objetivos, porém invasivos e caros). A análise acústica da voz busca extrair medidas robustas para descrever vários fenômenos associados à produção da fala ou características intrínsecas do ser humano como frequência fundamental, timbre, etc. O presente estudo consiste na caracterização de um modelo de processamento digital de Voz para apoio ao diagnóstico no contexto da construção de sistemas de identificação automatizados de patologias da fala. Para análise da técnica proposta foi utilizada uma base de dados (base KAY) que foi estruturada por especialistas num arranjo de seis grupos de Patologias. A esse, acrescentado também um de vozes “Normal”. Assim, 182 vozes foram escolhidas, as quais dispunham de um catálogo indexado de cerca de 33 descritores, para cada voz, calculados da elocução da vogal \a\ sustentada. Ao selecionar combinações desses descritores – como perturbações em frequência (jitter), em amplitude (shimmer) etc, este estudo encontrou evidências estatísticas e mostrou ser possível: a) Separar vozes normais das patológicas – esperado, b) Separar patologias específicas (Paralisia, Edema de Reinke, Nódulos) com acurácia de 100% (para a grande maioria dessas combinações) e cerca de 92% (para Nódulos contra Reinke); c) Discriminá-las por meio de classificadores (redes neurais artificiais e máquina de vetores de suporte) e reduzir a dimensionalidade e complexidade (quantidade de dados) via técnica de análise de componentes principais (ACP) sobre esses descritores para a separação intra patologias; e d) Testes estatísticos com os grupos locais confirmaram também limiares de indícios de Anormalidade presentes na literatura. A utilização de menor quantidade de descritores – obtida pós ACP (compressão) – mostrou-se também eficiente (mesmas taxas de acurácia). / It is estimated one-third of the work force relies on the use the voice in their jobs. The clinical diagnostic may be performed on voice listening by a specialist (subjective perspective) or through invasive and often not cheaper exams to check vocal structures. The area of Voice Acoustic analyses aims to extract robust measurements to describe several phenomena associated with voice production, or human being particular characteristics like fundamental frequency, timbre, etc. This study consisted of a model characterizing the digital voice processing for support in building automatic systems for the identification of disorders of speech (to aid diagnosis of pathologies). To support this investigation and proposed model, a commercial voice database (KAY base) was used with the endorsement from medical specialists. Derived acoustic analyses of those speech samples data records were presented to professionals for classification and six “severities groups” case-studied were built. After these analyses, one Normal group was added and, at the end, 182 voices have been selected. Their refined audio database contain, among other things, an indexed list of vocal descriptors calculated on the presence of the utterance of the vowel \a\ sustained speech. Statistical evidences were found: a) Difference between pathological groups vocal descriptors to normal (expected); b) It was achieved 100% from true positive, most cases, among Paralysis, Reinke's Edema and Nodules separations; c) from few cases, there were detected minor distinctions: Paralysis, Reinke's Edema, Nodules and Edema (pair comparison) with disordered groups; c) Among Machine Learning Algorithms (artificial neural networks "RN" and support vector machine "SVM"), the technique of Principal Components Analyses (PCA) and main statistics performed, it was found facts to help to structure some automated recognition systems. These Supervised learning methods showed that it could be possible to generate classification predictions (disordered presence) for the response to new data; and d) Inner tests also confirmed literature established reference thresholds. Hence considering suitable combinations of descriptors with two machine learning classifiers, as showed, is sufficient suitable and worthy.