The Recognition Component of the SUNDIAL Project

Abstract

The recognition component of the SUNDIAL system has being developed jointly by Logica Cambridge, Erlangen University, Cselt, Daimler Bern Ulm, Cap-Gemini Innovation and Politecnico di Torino; the device acts as an Acoustic Front-End, performing the feature extraction and the acoustic-phonetic decoding stages. For the feature extraction stage, several speech processing algorithms were tested and compared by means of RSA (Recognition Sensitivity Analysis) [13], in terms both of recognition performance and speaker-sensitivity. The recogniser was intended to be used over the telephone network: therefore, the problem of high dynamic range and of spectral drifts of the signal were addressed. To this purpose, energy and cepstrum normalisation procedures were introduced to improve robustness in real TLC environments. The acoustic-phonetic stage was based on the HMM technology with Discrete (Italian), Continuous (All languages) and Semi-Continuous (German) paradigms. Speech units were selected on a phoneme basis with context dependency. Tests on the SUNDIAL recogniser have been performed both with read and spontaneous speech. Recognition performance for a typical continuous speech, speaker independent task, based on read input, scored about 80% Word Accuracy over a 1000 words vocabulary, without linguistic constraints. The recognition component has been embedded in real-time prototypes built on the overall SUNDIAL architecture. This also includes linguistic processing, dialogue managing, access to the information system, message generation and text-to-speech synthesis functions. Four prototypes are also being tested with spontaneous dialogues obtained from naive speakers, in two different application environments: access to flight enquiry and reservation in English and French, and train information access in German and Italian.