Search

This paper describes the TALP participation in the WMT13 evaluation campaign. Our participation is based on the combination of several statistical machine translation systems: based on standard hrasebased Moses systems. Variations include techniques such as morphology generation, training sentence filtering, and domain adaptation through unit derivation. The results show a coherent improvement on TER, METEOR, NIST, and BLEU scores when compared to our baseline system.

This paper gives a description of the statistical machine translation (SMT) systems developed at the TALP Research Center of the UPC (Universitat Polit`ecnica de Catalunya) for our participation in the IWSLT’08 evaluation campaign. We present Ngram-based (TALPtuples) and phrase-based (TALPphrases) SMT systems. The paper explains the 2008 systems’ architecture and outlines translation schemes we have used, mainly focusing on the new techniques that are challenged to improve speech-to-speech translation quality. The novelties we have introduced are: improved reordering method, linear combination of translation and reordering models and new technique dealing with punctuation marks insertion for a phrase-based SMT system. This year we focus on the Arabic-English, Chinese-Spanish and pivot Chinese-(English)-Spanish translation tasks.

El comportamiento de los sistemas actuales de reconocimiento del habla se degrada rápidamente en presencia de ruido de fondo. Recientemente, se ha propuesto una técnica de representación de la señal de voz basada en la predicción lineal de la parte causal de la autocorrelación (OSALPC) que ha mostrado ser atractiva para el reconocimiento de habla ruidosa debido a sus altas prestaciones con respecto a la predicción lineal (LPC) convencional en condiciones severas de ruido blanco aditivo y a su simplicidad computacional. El propósito de este artículo es doble: 1) mostrar que la técnica OSALPC obtiene también buenas prestaciones en un entorno ruidoso real (ruido de coche), y 2) explorar su combinación con varias técnicas robustas de medida de similitud, mostrando que sus prestaciones mejoran aún más filtrando convenientemente los parámetros espectrales y realizando un etiquetado múltiple de los mismos. | The performance of the existing speech recognition systems degrades rapidly in the presence of background noise. A novel representation of the speech signal, which is based on Linear Prediction of the One-Sided Autocorrelation sequence (OSALPC), has shown to be attractive to speech recognition because of both its high recognition performance with respect to the standard LPC in severe conditions of additive white noise and its computational simplicity. The aim of this work is twofold: 1) to show that OSALPC also achieves good performance in a case of real noisy speech (in a car environment), and 2) to explore its combination with several robust similarity measuring techniques, showing that its performance even improves by filtering and multilabeling conveniently the spectral parameters.

Conventional modelling techniques of speech suffer a very big performance degradation in adverse noisy environments. So, it is necessary to research for more robust representations of speech signal. This paper presents new models that have succeeded in adverse environments. They are hybrid models of the classical parametrizations techniques used so far that have demonstrated being very useful in order to obtain good results in different noisy environments. In order to prove the their performance we have used white and machine noise in our experiments.

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 Quantization) approaches for speech recognition in noisy car environment: parameterization based on the linear prediction of the causal part of the autocorrelation sequence (OSALPC) - proposed by the authors in [1] [2]-, optimization of spectral model order and cepstral lifter, cepstral projection distance measure, dynamic information and multilabeling.

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 propose the application of a simple multilabeling method, instead of the standard vector quantization -so called labeling-, as the front end for a speech recognizer based on the Vector Quantization (VQ) and Hidden Markov Models (HMM) approaches in order to increase its robustness to noise. Furthermore, not only cepstrum but also other features such as energy and dynamic parameters are evaluated and quantized independently in the multilabeling stage to represent more accurately characteristics of speech. The result of this process is a multiple multilabeling. Experimental results in the presence of additive white noise clearly demonstrate its good performance in isolated word recognition in noisy environments.

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 propose the application of a simple multilabeling method, instead of the standard vector quantization -so called labeling-, as the front end for a speech recognizer based on the Vector Quantization (VQ) and Hidden Markov Models (HMM) approaches in order to increase its robustness to noise. Furthermore, not only cepstrum but also other features such as energy and dynamic parameters are evaluated and quantized independently in the multilabeling stage to represent more accurately characteristics of speech. The result of this process is a multiple multilabeling. Experimental results in the presence of additive white noise and car environment clearly demonstrate its good performance in isolated word recognition in noisy environments.

In the paper we present the theoretical development of the normalized backpropagation, and we compare it with other algorithms that have been presented in the literature. The algorithm that we propose is based on the idea of normalizing the adaptation step in the gradient search by the variance of the input. This algorithm is simple and gives good results in comparison with other algorithms that accelerate the learning and has the additional advantage that the parameters are calculated by the algorithm, so the user does not have to make several trials in order to trim the adaptation step and the momentum until the best combination is found. The task which we have designed in order to compare the algorithms is the recognition of digits in the Catalan language, with a data base of 1000 items, spoken by 10 speakers. The algorithms that we have compared with the normalized back propagation are: D.E.Rumelhart and J .L. McCielland, Franzini, Suddhard, Fahlman, Monte.

In this paper we investigate the use of a feature selection step in a isolated word recognition system. The feature selection step tries lo modal the correlalion among dajacent feature vectors and the variabilily of the speech. Thus, the feature selection is performed in two steps. The first step takes into account the temporal correlation among feature vectors in order to obtain a new set of feature vectors which are uncorrelated. This step gives a new template of M feature vectors, being M<

This paper describes a new coding structure based on the combination of Vector Quantizati.on, Linear Prediction l)nd Subband Splitting that achieves high guality speech at rates below 10 Kbit/sec. In this scheme, a vector is formed with one sanple of the normalized prediction error of each band and then a vector quanti.zer is applied to it. This guantization of the prediction error a.llows to use scalar adaptive predictors while conserving the advantages of the vector guantization. The necessary noise shap~ for achjev:ing high subjective quality is obtained by the use of a Freguency-Weighted distance in thc vecto r guantizer

The combination of time-domain harmonic scaling (TDHS) and sub-band coding (SBC) provides an encoding approach which allows 9.6 Kb/s speech encoding with good communication quality. Starting from this structure, this paper focuses the improvement of earlier designs. It is shown that adaptive prediction and bit-assigment enhances the subband signal coding and, hence, the performance of the overall system. The prediction is realized by an adaptive lattice, the algorithm being GAL2. The dynamic bit allocation takes place from the step-sizes of the backward adaptive quantizers (Jayant) in each sub-band. Improvements as high as 5 dB can be achieved for the average segmented signal to noise ratio.

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.

In this paper the voice quality of several 32-9.6 Kbit/s coders is assesed by a subjective test and by objective measures. A subjective test called Degradation Pairs Test (TDP) has been designed and the performance of objective measures in predicting the subjective quality is studied.

A vectorial generalization of the ADPCM system is introduced. Once the speech signal is grouped in vectors, they are coded using a vector predictor (VP) and a vector quantizer (VQ). Both subsystems are continously adaptive; the VQ is matched the signal power; two algorithms (VLMS and VGAL) are considered to adapt the predictor. At the bit stream corresponding to 2 bits per signal sample, the coding system rends a segmented signal to noise ratio equivalent to other existing coders.

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.