Search

Some Speech Enhancement algorithms based on the iterative Wiener filtering Method due to L1m-Oppenheim [2] are presented. In the original Lim-Oppenheim algorithm, speech AR estimation is carried out using classic second-order analysis, but our algorithms consider a more robust AR modelling. Two different strategies of speech AR estimation are presented and both estimators are trying to see as less amount of noise as possible. First one uses a previous One-Sided Autocorrelation computation, that is a pole-preserving function, and the actual SNR m the second-order LPC analysis is increased. Second one combines advantages of Higher-Order Statistics [1] with a linear combination of AR coefficients, belonging to two consecutive overlapped frames, to assess a less disturbed speech estimation.

In previous works [5], [6], we studied some speech enhancement algorithms based on the iterative Wiener filtering method due to Lim-Oppenheim [2], where the AR spectral estimation of the speech is carried out using a second-order analysis. But in our algorithms we consider an AR estimation by means of cumulant analysis. This work extends some preceding papers due to the authors: a cumulant-based Wiener Filtering (AR3_IF) is applied to Robust Speech Recognition. A low complexity approach of this algorithm is tested in presence of bathroom water noise and its performance is compared to classical Spectral Subtraction method. Some results are presented when training task of the speech recognition system (HTK-MFCC) is executed under clean and noisy conditions. These results show a lower sensitivity to the presence of water noise when applying AR3_IF algorithm inside of a speech recognition task.

We study some speech enhancement algorithms based on the iterative Wiener filtering method due to Lim-Oppenheim [2], where the AR spectral estimation of the speech is carried out using a second-order analysis. But in our algorithms we consider an AR estimation by means of cumulant analysis. This work extends some preceding papers due to the authors, where information of previous speech frames is taken to initiate speech AR modelling of the current frame. Two parameters are introduced to dessign Wiener filter at first iteration of this iterative algorithm. These parameters are the Interframe Factor (IF) and the Previous Frame Iteration (PFI). A detailed study of them shows they allow a very important noise suppression after processing only first iteration of this algorithm, without any appreciable increase of distortion. Finally, the simplest cumulant-based algorithm is applied to Speech Recognition and some preliminary results are presented.

The combination of Vector Quantization (VQ) and predictive techniques -named AVPC systems- has been shown as interesting systems for speech waveform coding at medium-high rates (ref. 1 to 3). In this work we present a such system including a previous four subbbands splitting. This allows good quality of speech at low-medium rates (1-1.25 bits/sample). A comparative study show that the AVPC-SBC outperforms the simple AVPC coder.

[ANGLÈS] In this project a musical tone classifier based on a neurally inspired onset model is performed, as many physiological and psychoacoustic evidences reveal the importance of the sound onset in the Human Auditory System, specifically in the fields of coding and recognition of sounds. The human inner ear is simulated: the cochlea is performed with a digital filterbank, the inner hair cells' behaviour is modelled with depressing synapses, and the auditory nerve is emulated with Leaky Integrate-and-Fire (LIF) neurons. After that, the onset spike trains obtained are transformed into a description matrix called onset fingerprint. The database used to test the performance of the system consists in 1020 single-note tones performed with five different instrument's families. The onset fingerprints of all these sounds are used to feed a classification system. Three approaches are made: Classification Trees, Quadratic Discriminant Classifier and Neural Networks. The last one gives a mean success classification rate of 75\%, the same performance as other onset-based classifier methods, but consuming less computing time. [CASTELLÀ] En este proyecto se realiza un clasificador de tonos musicales basado en el modelado del inicio del sonido (onset) inspirado en el comportamiento del nervio auditivo humano, ya que muchas evidencias fisiológicas y psicoacústicas ponen de manifiesto la importancia del inicio del sonido en los procesos de codificación y reconocimiento de sonidos del sistema auditivo humano. Con este propósito, se simula el oído interno humano: la cóclea es realizada con un banco de filtros digitales, el comportamiento de las células ciliadas internas se modela con sinapsis depresivas, y el nervio auditivo es emulado utilizando neuronas integradoras con pérdidas (leaky integrate-and-fire neurons). Una vez localizados temporalmente los inicios del sonido para diferentes canales y niveles de amplitud, se construye una matriz de descripción, llamada huella de inicio (onset fingerprint). La base de datos utilizada para evaluar el rendimiento del sistema consta de 1020 tonos realizados con cinco familias de instrumentos musicales diferentes. Las huellas de inicio, calculadas para todos estos sonidos, se utilizan para entrenar y testear tres sistemas de clasificación: árboles de clasificación, análisis discriminante cuadrático y redes neuronales. Este último da una tasa media de acierto del 75%, el mismo rendimiento que otros métodos basados ​​en esta aproximación biológica, pero usando menos tiempo de proceso. [CATALÀ] En aquest projecte es realitza un classificador de tons musicals basat en el modelatge de l'inici del so (onset) inspirat en el comportament del nervi auditiu humà, ja que moltes evidències fisiològiques i psicoacústiques posen de manifest la importància de l'inici del so en els processos de codificació i reconeixement de sons del sistema auditiu humà. Amb aquest propòsit, es simula l'oïda interna dels humans: la còclea és realitzada amb un banc de filtres digitals, el comportament de les cèl·lules ciliades internes es modela amb sinapsis depressives, i el nervi auditiu és emulat utilitzant neurones integradores amb pèrdues (leaky integrate-and-fire neurons). Un cop localitzats temporalment els inicis del so per a diferents canals i nivells d'amplitud, es construeix una matriu de descripció, anomenada empremta d'inici (onset fingerprint). La base de dades utilitzada per avaluar el rendiment del sistema consta de 1020 tons realitzats amb cinc famílies d'instruments musicals diferents. Les empremtes d'inici, calculades per a tots aquests sons, s'utilitzen per entrenar i testejar tres sistemes de classificació: arbres de classificació, anàlisi discriminant quadràtica i xarxes neuronals. Aquest últim dóna una taxa mitjana d'encert del 75%, el mateix rendiment que altres mètodes basats en aquesta aproximació biològica, però emprant menys temps de procés.