Search

Article publicat a Arxiv, sense revisió per parells In the analysis of acoustic scenes, often the occurring sounds have to be detected in time, recognized, and localized in space. Usually, each of these tasks is done separately. In this paper, a model-based approach to jointly carry them out for the case of multiple simultaneous sources is presented and tested. The recognized event classes and their respective room positions are obtained with a single system that maximizes the combination of a large set of scores, each one resulting from a different acoustic event model and a different beamformer output signal, which comes from one of several arbitrarily-located small microphone arrays. By using a two-step method, the experimental work for a specific scenario consisting of meeting-room acoustic events, either isolated or overlapped with speech, is reported. Tests carried out with two datasets show the advantage of the proposed approach with respect to some usual techniques, and that the inclusion of estimated priors brings a further performance improvement.

If speech is captured by several arbitrarily-located microphones in a room, the degree of distortion by noise and reverberation may vary strongly from one channel to another. Channel selection for automatic speech recognition aims to rank the signals according to their quality, and, in particular, to select the best one for further processing in the recognition system. To create this ranking, we propose here to use posterior probabilities estimated from the N-best hypothesis of each channel. When evaluated experimentally, this new channel selection technique outperforms the methods published so far. We also propose the combination of different channel selection techniques to further increase the recognition accuracy and to reduce the computational load without significant performance loss.

The performance of ASR systems in a room environment with distant microphones is strongly affected by reverberation. As the degree of signal distortion varies among acoustic channels (i.e. microphones), the recognition accuracy can benefit from a proper channel selection. In this paper, we experimentally show that there exists a large margin for WER reduction by channel selection, and discuss several possible methods which do not require any a-priori classification. Moreover, by using a LVCSR task, a significant WER reduction is shown with a simple technique which uses a measure computed from the sub-band time envelope of the various microphone signals.

El projecte es desenvolupa en el marc de la titulació oficial de màster MERIT del Departament de Teoria del Senyal i Comunicacions. El màster, orientat a la recerca i integrat dins del programa Erasmus Mundus, presenta trets esfecífics donat l’origen variat dels estudiants i que s’imparteix integrament en anglès. El projecte s’articula en 4 eixos: 1. Millorar el sistema d’acollida i integració dels estudiants en el programa formatiu del màster mitjançant la creació de recursos docents en xarxa (campus digital), i la incorporació de nous elements tecnològics a l’aula. 2. Millorar el seguiment dels estudiants de primer curs mitjançant suport als professors per part d’estudiants avantatjats de segon curs. 3. Fomentar el desenvolupament de competències transversals dels estudiants mitjançant la impartició d’assignatures que fomentin l’esperit crític i emprenedor. 4. Establir un mecanisme d’assegurament de la qualitat basat en un Comitè d’Avaluació Extern (CAE) d’acord amb les directrius de l’AQU

When applied to interactive seminars, the detection of acoustic events from only audio information shows a large amount of errors, which are mostly due to the temporal overlaps of sounds. Video signals may be a useful additional source of information to cope with that problem for particular events. In this work, we aim at improving the detection of steps by using two audio-based Acoustic Event Detection (AED) systems, with SVM and HMM, and a video-based AED system, which employs the output of a 3D video tracking algorithm. The fuzzy integral is used to fuse the outputs of the three detection systems. Experimental results using the CLEAR 2007 evaluation data show that video information can be successfully used to improve the results of audio-based AED.

A correct choice of voice signal modeling methods is essential to obtain good results in automatic speech recognition. In this paper, we have proposed a unified view of the speech parametrization stage, in which conventional techniques as Linear Prediction Coeficientes and mel-cepstrum filter bank are viewed as particular cases. The model incorporates a new technique of deconvolution, that is called root homomorphic deconvolution. A broad set of experimental results are also presented

The time sequences of speech parameters resulting from current short-time spectral estimators show a tradeoff between estimation error variance and time and frequency resolution. In this paper, we apply frequency analysis and linear filtering to these sequences to gain insights into their limitations and to provide an interpretation framework for several parameter processing techniques proposed in the past. Particularly, the observation of their long-term spectrum reveals the importance of band equalization for improving discrimination in speech recognition. Based on that, we propose a method of filtering the sequences that includes an explicit equalization and incorporates a bandwidth parameter. By using Slepian sequences in the design of the filters, good results were obtained in our preliminary word recognition tests.

The performance of existing speech recognition systems degrades rapidly in the presence of background noise when training and testing cannot be done under the same ambient conditions. The aim of this paper is to report the application of several robust techniques on a system based on the HMM (Hidden Markov Models) and VQ (Vector Quatization) approaches for speech recognition in car environment optimization of spectral model order and cepstral lifter, parameterizatrion based on the AR modeling of the autocorrelation sequence -proposed by the authors in [1]-, use of dynamic information and multilabeling.

Concepts and techniques involved in speech communication with computers are reviewed. Although the speech output is, at present, closer to the computer ability than the speech understanding, at the moment only partial results for both purposes have been attained.

Article publicat sense revisió per parells a Arxiv In the noisy acoustic environment of a Neonatal Intensive Care Unit (NICU) there is a variety of alarms, which are frequently triggered by the biomedical equipment. In this paper different approaches for automatic detection of those sound alarms are presented and compared: 1) a non-model-based approach that employs signal processing techniques; 2) a model-based approach based on neural networks; and 3) an approach that combines both non-model and model-based approaches. The performance of the developed detection systems that follow each of those approaches is assessed, analysed and compared both at the frame level and at the event level by using an audio database recorded in a real-world hospital environment.

In this paper we study various decorrelation methods for the features used in speech recognition and we compare the performance of each one by running several tests with a speech database. First of all we study the Principal Components Analysis (PCA). PCA extracts the dimensions along which the data vary the most, and thus it allows us to reduce the dimension of the data point without significant loss of performance. The second transform we study is the Discrete Cosine Transform (DCT). As it will be shown, it is an approximation of the PCA analysis. By applying this transform to FBE parameters we obtain the MFCC coeficients. A further step is taken with the Linear Discriminant Analysis (LDA), which, not only reduces the dimensionality of the problem, but also discriminates among classes to reduce the confusion error. The last method we study is Frequency Filtering (FF). This method consists of a linear filtering of the frequency sequence of the log FBE that both decorrelates and equalizes the variance of the coeficients.

The analysis of acoustic scenes requires several functionalities, being perhaps recognition (speech, speaker, other acoustic events) and spatial localization the two most relevant ones. For a reduced invasiveness, the microphones are far away from the sound sources, and possibly grouped in arrays, which may be distributed, not arranged, in the room. Aiming at an increased performance, the usual model- based approach employed for sound recognition or detection can be extended to other co-occurrent tasks like source localization, so both tasks can be carried out jointly, using the same formulation and processing. In this paper, we intend to illustrate that point by presenting together a few new model-based techniques that deal with the problems of overlapped-sounds recognition, multi-source localization, and channel selection. They are briefly described, and tested in a smart-room environment with a multiple microphone- array setup.

Cepstral coefficients obtained by means of a linear predictive analysis (LPC) of either the signal or a related sequence have given remarkable robustness against additive Gaussian white noise. From a comparative study of the most robust techniques appeared in the literature and some shight alternatives proposed in this paper, we attain two main conclusions : 1) the slope cepstral window and a relatively high model order are preferable; and 2) the spectral representation based on the autocorrelation (rather than on the signal) modelling achieves better results than the high performance cepstral projection distance measure for low signal-to-noise ratios.

En esta comunicación se describe el uso de la semisílaba en el reconocimiento de habla continua para una aplicación específica: el reconocimiento de los números enteros castellanos del cero al mil. Tras una breve descripción de la arquitectura del sistema de reconocimiento, se detalla la inferencia de la gramática de estados finitos que representa los números en términos de semisílabas, y se indica el procedimiento seguido para la generación de las referencias de las mismas. Finalmente, se presentan los resultados alcanzados en dos experimentos: en el primero el sistema de reconocimiento es entrenado para un locutor y las referencias utilizadas para las semisílabas son patrones de características frecuenciales; en el segundo, el entrenamiento es multilocutor y las semisílabas son representadas mediante Modelos Ocultos de Markov. En ambos casos la tasa de reconocimiento del sistema es excelente.

Cepstral coefficients are widely used in speech recognition. In this paper, we claim that they are not the best way of representing the spectral envelope, at least for some usual speech recognition systems. In fact, cepstrum has several disadvantages: poor physical meaning, need of transformation, and low capacity of adaptation to some recognition systems. In this paper, we propose a new representation that significantly outperforms both mel-cepstrum and LPC-cepstrum techniques in both recognition rate and computational cost. It consists of filtering the frequency sequence of filter-bank energies with an extremely simple filter that equalizes the variance of the cepstral coefficients. Excellent results of the new technique using a continuous observation density HMM recognition system and two very different recognition tasks, connected digits and phone recognition, are presented.

La recerca i el desenvolupament d'aplicacions en tecnologies de la parla han deixat força de banda els coneixements lingüístics. Les raons són diverses, com veurem, però la dificultat del treball conjunt de lingüistes i tecnòlegs no significa que aquest no sigui necessari per arribar més lluny en els objectius de la pròpia tecnologia.

Real-time processing is a requirement for many practical signal processing applications. In this work we implemented online 2-source acoustic event detection and localization algorithms in a Smart-room, a closed space equipped with multiple microphones. Acoustic event detection is based on HMMs that enable to process the input audio signal with very low latency; acoustic source localization is based on the SRP-PHAT localization method which is known to per-form robustly in most scenarios. The experimental results from online tests show high recognition accuracy for most of acoustic events both isolated and overlapped with speech.

Recently, the set of spectral parameters of every speech frame that result from filtering the frequency sequence of mel-scaled filter-bank energies with a simple first-order high-pass FIR filter have proved to be an efficient speech representation in terms of both speech recognition rate and computational load. In this paper, we apply the same technique to speaker recognition. Frequency filtering approximately equalizes the cepstrum variance, enhancing the oscillations of the spectral envelope curve that are most effective for discriminating between speakers. In this way, even better speaker identification results than using conventional mel-cepstrum were observed in continuous observation Gaussian density HMM, especially in noisy conditions.

SIMBAD has been developed to facilitate the tasks involved in the design of a concatenative speech synthesis system: 1) building a dictionary of parameterized speech units, and 2) obtaining a set of rules to concatenate these units and to model prosodics. It is an interactive, menu-driven and graphical software tool that allows both automatic processing and graphical editing of speech with high flexibility.

This paper describes the phonetic content of Albayzin, a spoken database for Spanish designed for speech recognition purposes. A statistical study of a large sample of spontaneous speech is presented, and the phonetic and statistical criteria for the final constitution of the database are discussed. Finally, the contents of the phonetic database are analyzed

In this work, we show how speaker-independent CDHMM word recognition performance can be significantly improved for clean speech by filtering the time sequence of spectral parameters to enhance its time dynamics. Experimental results with the standard TI connected digits database show the filter can achieve more than 30% reduction of string recognition error. As shown in this paper, that improvement is partially due to the speaker variability reduction obtained by attenuating the very low modulation frequencies. The widely used cepstral mean subtraction technique also improves the recognition rate, but it can not achieve such a noticeable improvement as the parameter filter. In fact, the best results are obtained when the peak of the long-term spectrum of the filter output is at around 3 Hz, a frequency which corresponds to the average syllable rate of the employed database.