Train times? Yes, ask the machine

Robots, machines that speak, answering machines that understand what we say … will be soon a regular part of our daily life. Concretely the University of the Basque Country (EHU/UPV), together with the universities of Zaragoza and Valencia, is developing a system capable of recognising speech. The aim of the project is to develop a machine which responds automatically to the user who asks for information about trains and timetables.

The machine will be able to recognise the voice of the person asking for information, understand what is being said and then find the data and respond appropriately. In order to carry this function out, the machine will have to synthesise the voice.

Speech recognition

The first step the machine has to take is speech recognition. This function is being developed by the UPV/EHU research group which first has had to gather together all the phonemes in a language, in this case Spanish and to these are then added information about different manners of pronunciation and about coarticulation. It is well known that a phoneme is pronounced in a differentiated manner depending on what precedes or follows it – this is known as coarticulation. Moreover, the differences that can arise depending on context are also taken into account. Finally, with all this data, phoneme patterns are drawn up.

In a second phase, the pattern of the language is completed; the pattern which, in this context, is to be used. Given that the language register used in the questions is limited in breadth and depth – the language of aerospace, for example, is not going to be used – a person asking for information about trains uses a specific, limited register of words and expressions.

Subsequently, the phonemes and language patterns are unified and features of spontaneous speech are added; for example, the pauses that we make between one word and another, linguistic fillers or other sounds without meaning, repetitions and so on.

Once all these patterns are unified, the research team drew up an information programme which can carry out a statistic analysis of those questions and the ways of effecting them as represented in samples and interprets what the end-user is saying in each case. Thus, it can be said that understanding speech is the first step.

Subsequently, the system developed by UPV/EHU is integrated with the rest of the units involved: that of conversation management, the one that seeks the date asked for, the unit that creates the response, that which synthesises the voice and, finally, the unit that transmits the information. Moreover, these units are not located in one, single site but are spread out in Zaragoza, Valencia and Leioa (the Bilbao campus of the Basque University) and, thus, large amounts of information are sent in real time.

Because of this the system has to be secure and, moreover, having the units physically separated confers another advantage on the system: it is modular. So, if one of the units fails, it is only a matter of fixing or changing that unit.

Contact :
Garazi Andonegi
ELHUYAR Fundazioa
garazi@elhuyar.com
(+34) 943363040