Igor Odriozola

ML Parameter Generation with a Reformulated MGE Training Criterion - Participation in the Voice Conversion Challenge 2016.

This paper describes our entry to the Voice Conversion Challenge 2016. Based on the maximum likelihood parameter generation algorithm, the method is a reformulation of the minimum generation error training criterion. It uses a GMM for soft classification, a Mel-cepstral vocoder for acoustic analysis and an improved dynamic time warping procedure for source-target alignment. To compensate the oversmoothing effect, the generated parameters are filtered through a speaker-independent postfilter implemented as a linear transform in cepstral domain. The process is completed with mean and variance adaptation of the logfundamental frequency and duration modification by a constant factor. The results of the evaluation show that the proposed system achieves a high conversion accuracy in comparison with other systems, while its naturalness scores are intermediate.

Development and evaluation of AnHitz, a prototype of a basque-speaking virtual 3D expert on science and technology

The aim of the AnHitz project, whose participants are research groups with very different backgrounds, is to carry out research on language, speech and visual technologies for Basque. Several resources, tools and applications have been developed in AnHitz, but we have also integrated many of these into a prototype of a 3D virtual expert on science and technology. It includes Question Answering and Cross Lingual Information Retrieval systems in those areas. The interaction with the system is carried out in Basque (the results of the CLIR module that are not in Basque are translated through Machine Translation) and is speech-based (using Speech Synthesis and Automatic Speech Recognition). The prototype has received ample media coverage and has been greatly welcomed by Basque society. The system has been evaluated by 50 users who have completed a total of 300 tests, showing good performance and acceptance.

An on-line VAD based on Multi-Normalisation Scoring (MNS) of observation likelihoods

Abstract Voice activity detection (VAD) is an essential task in expert systems that rely on oral interfaces. The VAD module detects the presence of human speech and separates speech segments from silences and non-speech noises. The most popular current on-line VAD systems are based on adaptive parameters which seek to cope with varying channel and noise conditions. The main disadvantages of this approach are the need for some initialisation time to properly adjust the parameters to the incoming signal and uncertain performance in the case of poor estimation of the initial parameters. In this paper we propose a novel on-line VAD based only on previous training which does not introduce any delay. The technique is based on a strategy that we have called Multi-Normalisation Scoring (MNS). It consists of obtaining a vector of multiple observation likelihood scores from normalised mel-cepstral coefficients previously computed from different databases. A classifier is then used to label the incoming observation likelihood vector. Encouraging results have been obtained with a Multi-Layer Perceptron (MLP). This technique can generalise for unseen noise levels and types. A validation experiment with two current standard ITU-T VAD algorithms demonstrates the good performance of the method. Indeed, lower classification error rates are obtained for non-speech frames, while results for speech frames are similar.

The AhoSR Automatic Speech Recognition System

AhoSR is a hidden Markov model based speech recognition system developed in the Aholab Signal Processing Laboratory research group of the University of the Basque Country. It has been modularly devised for ASR-based tools and applications to be easily implemented and tested, being also particularly interesting for research in the field of language model optimization of agglutinative languages like Basque. The system relies on the use of a static search graph where decoupled language model information is incorporated at run-time. This paper introduces the basic architecture as well as the most relevant aspects of the AhoSR speech recognition system. Besides, this paper compiles the results of several experiments which validate the system for its use in different tasks: phonetic, grammar-based and LM-based recognition. Two CALL/CAPT applications that use AhoSR are also described.

Risk for our Languages and A Challenge for Language Technology

We are witnesses to a digital revolution that is dramatically impacting communication and society. Recent developments in digital information and communication technology are sometimes compared to Gutenberg’s invention of the printing press. What can this analogy tell us about the future of the European information society and our languages in particular?

Language Technology Support for Basque

Language technology is used to develop software systems designed to handle human language and are therefore often called “human language technology”. Human language comes in spoken and written forms. While speech is the oldest and in terms of human evolution the most natural form of language communication, complex information and most human knowledge is stored and transmitted through the written word. Speech and text technologies process or produce these different forms of language, using dictionaries, rules of grammar, and semantics. This means that language technology (LT) links language to various forms of knowledge, independently of the media (speech or text) in which it is expressed.

Arriskua, Gure Hizkuntzentzat Eta Erronka Hizkuntza-Teknologiarentzat

Komunikazioan eta gizartean izugarrizko eragina izaten ari den iraultza digital baten aurrean gaude. Komunikazio-teknologia digitalizatu eta sarekoetan izan berri diren aurrerapenak Gutenbergek inprenta asmatu zuenekoarekin alderatzen dira, batzuetan. Zer esaten digu analogia horrek Europako informaziogizartearen eta geure hizkuntzen etorkizunari buruz?

Hizkuntza-Teknologia Euskararako

Hizkuntza-teknologiak giza hizkuntzarekin lan egiteko espezializatutako informazio-teknologiak dira. Horregatik, giza hizkuntzaren teknologia izenpean ere ezagutzen dira maiz teknologia hauek. Giza hizkuntza ahoz nahiz idatziz agertzen da. Hizkuntza-komunikazioko modurik zaharrena eta naturalena hizketa bada ere, informazio konplexua eta giza ezagutzaren zati handiena testu idatzien bidez gorde eta transmititzen da. Hizketaeta testu-teknologiek bi modu horietan prozesatzen edo sortzen dute hizkuntza. Baina, hizkuntzak baditu hizketan nahiz testuetan agertzen diren alderdiak ere, hala nola hiztegiak, gramatikaren zati handi bat eta esaldien esanahia. Hortaz, hizkuntza-teknologiako atal asko ezin dira bietako batean sartu, hizketa-teknologian ala testu-teknologian. Horien artean daude hizkuntza ezagutzarekin lotzen duten teknologiak.

Euskara Europako Informazio Gizartean

Euskara, Nafarroako Erresumako hizkuntza nagusia zelako latinez “Lingua Navarrorum” esaten zitzaiona, mendebaldeko Europan bizirik dagoen hizkuntza preindoeuropar bakarra da. Hizkuntza bakartutzat jotzen da, ez baitzaio loturarik aurkitu beste hizkuntzekin, antzinako akitanierarekin izan ezik. Euskararen jatorria nahiz beste hizkuntzekiko duen lotura gai gatazkatsuak eta interesgarriak dira oraindik ikerlarientzat.

Meta-Neti Buruz

Europako Batzordeak sortutako bikaintasunezko sarea da META NET. Sareak Europako 33 herrialdetako 54 kide ditu, gaur egun. META-NETek META, Europa Eleaniztunaren Teknologia Aliantza, babesten du, hizkuntza-teknologiako aditu eta erakundeen talde europar gero eta handiagoa. META-NETek oinarri teknologikoak eman nahi ditu informazio-gizarte zinez eleaniztuna sortzeko Europan, eta hari eusteko.