Hans-Wilhelm Rühl

System with speaking-rate-adaptive transition values for determining words from a speech signal

Speech recognition produces test signals from the speech signal which are compared with predetermined reference signals so as to form scores. Each subsequent test signal is compared with reference values which are situated within a predetermined neighborhood of the reference value which has been determined to be optimum for the preceding test signal. In dependence on this neighborhood, transition values in conformity with the transition probabilities are added to the scores. In order to enhance the results notably in the case of different speeds of speaking of the instantaneous speaker, it is proposed to adapt these transition values in dependence on the speed of speaking. A further improvement can be achieved by also adapting the reference values themselves to the relevant speakers pronunciation. This adaptation can also be iteratively performed in a number of steps.

Evaluation of a TTS-System Intended for the Synthesis of Names

A German speech synthesis system is presented for telephone-based inquiry systems supplying address-oriented information. Providing speech synthesizers for public inquiry systems is a very demanding task, as most users of such systems expect a speech quality close to natural and perfect intelligibility. To achieve these goals, a synthesis-by-concept scheme was developed for speech output allowing one to combine synthetically spoken names with naturally spoken carrier sentences, enhanced by special rule sets for grapheme-phoneme conversion of first names, and family names, business names, street names, and location names.

A Server for Area Code Information Based on Speech Recognition and Synthesis by Concept

An automatic information server to be operated via the telephone net is presented, requiring the input of an old German area code for mail adresses and responding with the new area code. The server both employs speaker-independent recognition of connected words, and a synthesis scheme by concept speaking town names by rules. Recognition is based on an HMM-recognizer, trained Speaker independent. Its feature extraction is insensitive both to different frequency responses of telephone and transmission line, and to slowly varying background noise. Speech synthesis uses PSOLA technique and concatenation of diphones. In order to generate highly natural sounding speech, most utterances are stored as natural speech, with only town names being synthesized. To prevent the impression of two different voices, utterances and diphones were spoken by the same speaker.