Alberto Ciaramella

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.

Device for speaker's verification

The device obtains several characteristic parameters from a standard sentence said by a speaker and compares them with average parameters of the same speaker stored in an internal memory and previously calculated. According to the comparison, it obtains a probability value that the sentence spoken belongs to that speaker and compares the value with a threshold normalized to the average parameter variance by a threshold calculation circuit. If the threshold is overcome, the device considers the speaker verified. A circuit determines the real instants of sentence beginning and end using a noise-adaptive threshold in order to limit between these two instants the time interval over which characteristic parameters are to be calculated. A circuit aligns as to time the characteristic parameters just calculated to the parameters of a reference sentence, obtaining standard lengths of the sounds composing the sentence spoken. A variable probability threshold is controlled by the standard deviations of the histogram of the average of the characteristic parameter vectors.

Test and evaluation of a spoken dialogue system

The development of spoken dialogue systems (SDSs) requires the definition of evaluation metrics which can assess the performance of these systems at different levels and compare various SDSs. The authors present a first test, made with naive users, on an integrated dialogue system for telephone speech access to a remote data base. They describe the system architecture as well as the goals of the test, its features, the methodology used during the evaluation, and the results obtained. The SDS is shown to be effective for providing the user with the required information. The presence of spontaneous speech phenomena is frequent with naive users. The dialogue helps the user to overcome the errors due to spontaneous speech. The use of isolated words for confirmation is useful, but partially limits the interaction friendliness.<<ETX>>

Implementation of an acoustical front-end for speech recognition

We describe the implementation of a programmable general-purpose acoustical front-end for speech recognition; its design keeps into account, as an example, the algorithm of centisecond cepstrum extraction for an acoustical signal sampled at a maximum rate of 12.8 kHz.

Conclusions and Future Developments

The speech understanding system that has been described in the preceding chapters represents a first prototype that is still open to evolution and improvement. Although the final result of the ESPRIT Project P26 represented a good achievement of our goals, we do believe that both the recognition and understanding stage should get better performance in order to make the overall system more accurate and robust.

The Real Time Implementation of the Recognition Stage

Subtasks 2.2 and 2.3 of the P26 project have been devoted to the design of a hardware architecture and to the implementation on it, in real time, of recognition algorithms already developed and experimented within Subtask 2.1.: this real time implementation of the recognition stage will be called RICO in the following. Table 3.1 summarizes the key points we considered when we started our work, i.e. algorithmic requirements, project development constraints, hardware and software technology limits; they contributed to the definition of RICO main characteristics, summarized in Table 3.2: in the following of this paragraph we will detail these considerations. We started with the consideration that recognition algorithms can be distinguished into two principal blocks, a first “feature extraction” block till vector quantization and phonetic classification of frames, and a following “search” block extracting the lattice of most likely words using dynamic programming: this system “cut” corresponds to the minimal flow of data and besides separates blocks with different computational characteristics. The first block in fact is characterized by predictable execution times, cyclic computations, vector data structures, not-too-large data addressing requirements: this block in fact implements “traditional” DSP algorithms, for which the DSP chips fit well. Instead memory and computational requirements of the second block heavily depend on the recognition vocabulary size and on the speaking style (continuous speech of course is more demanding than isolated words) and also exhibit a time dependency for the same utterance; in each case, for the real time recognition of continuous speech with a 1K words vocabulary, the computational requirements are quite demanding, although were not clearly defined at the beginning of the project.

Vector quantization firmware for an acoustical front-end using the TMS32020

Here we describe the firmware implementation of an acoustical front-end, performing the vector quantization of Discrete Cosine Transform (DCT) for a speech recognition system. This firmware runs on a single TMS32020 signal processor chip and is characterized both by a substantial real time performance and by a good accuracy.