Jon Ander Gómez

Comparative study of automatic phone segmentation methods for TTS

We present two novel approaches to phonetic speech segmentation. One is based on acoustical clustering plus dynamic time warping and the other is based on a boundary specific correction by means of a decision tree. The use of objective or perceptual evaluations is discussed. The novel approaches clearly outperform the objective results of the baseline system based on HMM. They get results similar to agreement between manual segmentations. We show how phonetic features can be successfully used for boundary detection together with HMMs. Finally, the need for perceptual tests in order to evaluate segmentation systems is pointed out.

Improvements on automatic speech segmentation at the phonetic level

In this paper, we present some recent improvements in our automatic speech segmentation system, which only needs the speech signal and the phonetic sequence of each sentence of a corpus to be trained. It estimates a GMM by using all the sentences of the training subcorpus, where each Gaussian distribution represents an acoustic class, which probability densities are combined with a set of conditional probabilities in order to estimate the probability densities of the states of each phonetic unit. The initial values of the conditional probabilities are obtained by using a segmentation of each sentence assigning the same number of frames to each phonetic unit. A DTW algorithm fixes the phonetic boundaries using the known phonetic sequence. This DTW is a step inside an iterative process which aims to segment the corpus and re-estimate the conditional probabilities. The results presented here demonstrate that the system has a good capacity to learn how to identify the phonetic boundaries.

Automatic speech segmentation based on acoustical clustering

In this paper, we present an automatic speech segmentation system based on acoustical clustering plus dynamic time warping. Our system operates at three stages, the first one obtains a coarse segmentation as a starting point to the second one. The second stage fixes phoneme boundaries in an iterative process of progressive refinement. The third stage makes a finer adjustment by considering some acoustic parameters estimated at a higher subsampling rate around the boundary to be adjusted. No manually segmented utterances are used in any stage. The results presented here demonstrate a good learning capability of the system, which only uses the phonetic transcription of each utterance. Our approach obtains similar results than the ones reported by previous related works on TIMIT database.

A Phonetic-Based Approach to Query-by-Example Spoken Term Detection

Query-by-Example Spoken Term Detection QbE-STD tasks are usually addressed by representing speech signals as a sequence of feature vectors by means of a parametrization step, and then using a pattern matching technique to find the candidate detections. In this paper, we propose a phoneme-based approach in which the acoustic frames are first converted into vectors representing the a posteriori probabilities for every phoneme. This strategy is specially useful when the language of the task is a priori known. Then, we show how this representation can be used for QbE-STD using both a Segmental Dynamic Time Warping algorithm and a graph-based method. The proposed approach has been evaluated with a QbE-STD task in Spanish, and the results show that it can be an adequate strategy for tackling this kind of problems.

End-to-end neural network architecture for fraud scoring in card payments

A completely new solution for fraud scoring in card payments validated by one of the biggest bank in the world.Experiments were performed with real dataset with more than 900 millions of credit card operations.Experiments performance is comparable to state of the art solutions but without any human intervention. Millions of euros are lost every year due to fraudulent card transactions. The design and implementation of efficient fraud detection methods is mandatory to minimize such losses. In this paper, we present a neural network based system for fraud detection in banking systems. We use a real world dataset, and describe an end-to-end solution from the practitioners perspective, by focusing on the following crucial aspects: unbalancedness, data processing and cost metric evaluation. Our analysis shows that the proposed solution achieves comparable performance values with state-of-the-art proprietary and costly solutions.

PRHLT: Combination of Deep Autoencoders with Classification and Regression Techniques for SemEval-2015 Task 11

This paper presents the system we developed for Task 11 of SemEval 2015. Our system had two stages: The first one was based on deep autoencoders for extracting features to compactly represent tweets. The next stage consisted of a classifier or a regression function for estimating the polarity value assigned to a given tweet. We tested several techniques in order to choose the ones with the highest accuracy. Finally, three regression techniques revealed as the best ones for assigning the polarity value to tweets. We presented six runs corresponding to three regression different techniques in combination with two variants of the autoencoder, one with input as bags of words and another with input as bags of character 3grams.

Using Word Graphs as Intermediate Representation of Uttered Sentences

We present an algorithm for building graphs of words as an intermediate representation of uttered sentences. No language model is used. The input data for the algorithm are the pronunciation lexicon organized as a tree and the sequence of acoustic frames. The transition between consecutive units are considered as additional units.