Luciano Nebbia

Interactive voice technology at work: the CSELT experience

Abstract This paper is a survey of the speech technologies and applications developed at CSELT, some of which are employed in real services deployed in the Italian telephone network. With the rise of significant speech recognition and text-to-speech applications, the activity of our lab encompasses now a broader set of activities, from new algorithmic approaches to speech product engineering and application development. In particular, the paper gives an overview of the products originated from our speech technology research. It describes two operative applications, namely a voice dialing service for large name directories, which is installed in the CSELT PABX, and an automated network service for directory assistance, which is now accessible to all the Italian telephone customers.

Eight-channel digital speech synthesizer based on LPC techniques

An automatic vocal response system for the Italian language has been implemented at CSELT, consisting of a hardware speech synthesizer controlled by a programmed device (mini or micro computer). The synthesizer exploits a speech production model composed of a 10th order digital lattice filter and two excitation generators for voiced and unvoiced sounds. The hardware includes also a module, which controls the updating and transfer of the parameters, and an output module which provides the analog speech signal. The synthesizer configuration is modular and expandible up to 8 channels. For each channel, the minicomputer supplies the synthesizer with the start-stop command plus 13 parameters: 10 filter coefficients, a gain factor, the pitch period and voiced-unvoiced information and the updating interval. For each channel, every 125 µs, 20 multiplications, 9 addition and 10 subtractions are executed. The filter and the source generator are time-shared among the 8 channels. The complete digital equipment is implemented by TTL-LS integrated circuits.

A specialised speech synthesis technique for application to automatic reverse directory service

This paper describes a specialised version of Eloquens, the CSELTs text-to-speech synthesiser, which has been conceived to achieve a substantial improvement for what concerns not only intelligibility, but also speech naturalness in the synthesis of messages in an automatic reverse directory service. This result has been obtained taking advantage of the peculiarities of the application domain, namely the possibility, or even the preference, to synthesise only isolated words and the restriction of the lexical domain to words occurring in the telephone directory. The system is based on plain concatenative synthesis, using acoustic units larger than diphones and avoiding prosodic manipulation. Coverage completeness is assured by the conventional diphones, which the system can use to synthesise possible missing units. Subjective evaluation demonstrates that the new system has higher intelligibility, requires less comprehension effort and shows a highly improved system acceptance than standard Eloquens.

Comparison of some algorithms for tap weight evaluation in adaptive echo cancellers

The performance of some adaptive algorithms for coefficient updating of a digital echo canceller are compared. The examined algorithms are the least mean square one (LMS), the normalized LMS and the simplified versions employing the sign information. Their performance are evaluated on the basis of the echo return loss enhancement (ERLE) steady state value and convergence speed. For the algorithms employing a linear function of the error, as gradient estimate, the resuIts show that convergence speed is dependent on the echo canceller tap number and that its trend is exponential. Algorithms employing the error sign as gradient estimate are the slowest if the same variance of residual echo must be obtained. Furthermore some consideration are made in comparison with an algorithm designed in order to minimize the mean square error evaluated over a M sample block.