Speech intelligibility represents how comprehensible a speech is. It is more important than speech quality in some applications. Single channel speech intelligibility enhancement is much more difficult than multi-channel intelligibility enhancement. It has recently been reported that training-based single channel speech intelligibility enhancement algorithms perform better than Signal to Noise Ratio (SNR) based algorithm. In this thesis, a training-based Deep Neural Network (DNN) is used to improve single channel speech intelligibility. To increase the performance of the DNN, the Multi-Resolution Cochlea Gram (MRCG) feature set is used as the input of the DNN. MATLAB objective test results show that the MRCG-DNN approach is more robust than a Gaussian Mixture Model (GMM) approach. The MRCG-DNN also works better than other DNN training algorithms. Various conditions such as different speakers, different noise conditions and reverberation were tested in the thesis.

Made available in DSpace on 2016-08-17T14:52:55Z (GMT). No. of bitstreams: 1 Natalia Costa Leite Abreu.pdf: 841490 bytes, checksum: 00ff55b62f0819b502a66a2304564bf4 (MD5) Previous issue date: 2003-12-11 / The constant search for the improvement and strengthening of the relationship between humans and machines turning it more natural is common place. Consequently, the recognition of speech will turn, easier and practical the handling of equipments supplied with the capacity to understand the human speech. In this sense and with the use of the available knowledge information in the literature as how the human brain processes informations, some suggested methods try to simulate this ability in the computer, especially devoted to the extraction of a speech signal of mixed sounds, attempting, for example to increase the recognition and comprehension rate. The extraction of speech can be obtained by measures of a single-channel or multiple the channels. In order to extract the speech in a single channel, it is proposed here to use the speech characteristics introducing the concept of efficient codification, that tries to imitate the way the auditory cortex gets information using the method of Independent Component Analysis (ICA), getting the basis functions of the input signals and retrieving the estimated signal even when we add interferences to it. Our simulations also prove the efficiency of our method against reverberation effects and the recovery of speech signal by the handling of basis function of other speech signals. This technique can be used efficiently both to extract a single speech, as well as highlighting new ways of approaching the speech/speaker recognition problem. / A constante busca para aperfeiçoar e estreitar o relacionamento entre homens e máquinas, tornando-o mais natural, não é nenhuma novidade. Conseqüentemente, o reconhecimento da voz possibilitará uma manipulação mais fácil e prática de equipamentos dotados com a capacidade de compreender a fala humana. Neste sentido e utilizando-se dos conhecimentos disponíveis na literatura de como o cérebro humano processa informações, alguns métodos propostos procuram simular computacionalmente essa habilidade, voltados principalmente à extração de um sinal de voz de uma mistura de sons, na tentativa de, por exemplo, aumentar a taxa de reconhecimento e inteligibilidade. A extração da voz pode ser obtida usando medidas de um único ou múltiplos canais. Para extrair uma voz em um único canal, propomos usar as características da voz pelo conceito de codificação eficiente, que procura imitar o modo como o córtex auditivo trata as informações, utilizando-se da técnica de Análise de Componentes Independentes (ICA), obtendo as funções bases dos sinais de entrada e recuperando o sinal estimado, mesmo quando são adicionadas interferências. Através de simulações comprovamos também a eficiência da técnica usada, primeiro, na recuperação de um sinal de voz com a utilização das funções bases de outro sinal e, segundo, frente a efeitos de reverberação. Esta técnica pode ser usada para extrair uma única fala eficazmente, como também prenuncia um modo novo de chegar ao problema de reconhecimento da fala/orador.