The requirements of the computer telephony (CT) industry place conflicting demands on text-to-speech (TTS) systems. Multilingual functionality and high quality output at telephone bandwidth requires detailed linguistic and acoustic analysis. At the same time, the need for robustness together with a high channel count and small memory footprint means that systems must be extremely efficient and databases must be kept small. We present a system which provides TTS for six languages, with 100 channels of highly natural output on a single DSP card.

One of the most successful approaches to synthesizing speech, concatenative synthesis, combines recorded speech units to build full utterances. However, the prosody of the stored units is often not consistent with that of the target utterance and must be altered. Furthermore, several types of mismatch can occur at unit boundaries and must be smoothed. Thus, both pitch and time-scale modification techniques as well as smoothing algorithms play a crucial role in such concatenation based systems. In this paper, we describe novel approaches to each of these issues. First, we present a conceptually simple technique for pitch and time-scale modification of speech. Our method is based upon a harmonic coding of each speech frame, and operates entirely within the original sinusoidal model. Crucially, it makes no use of "pitch pulse onset times." Instead, phase coherence, and thus shape invariance, is ensured by exploiting the harmonic relation existing between the sine waves used to code each analysis frame so that their phases at each synthesis frame boundary are consistent with those derived during analysis. Secondly, a smoothing algorithm, aimed specifically at correcting phase mismatches at unit boundaries, is described. Results are presented showing our prosodic modification techniques to be highly suitable for use within a concatenative speech synthesizer.

A new and simple approach to shape invariant time-scale modification of speech is presented. The method based upon a harmonic coding of each speech frame operates entirely within the original sinusoidal model and makes no use of "pitch-pulse onset times" used by conventional algorithms. Instead, phase coherence, and thus shape invariance, are ensured by exploiting the harmonic relation existing between the sine waves to cause them to be in phase at each adjusted frame boundary. Results suggest this approach to be an excellent candidate for use within a concatenative text-to-speech synthesiser where scaling factors typically lie within a range well handled by this algorithm.

The growing industrial take-up of spoken language dialogue systems (SLDSs), their constantly increasing sophistication, and the scarcity of teams which master the full system complexity as well as all the necessary steps in the SLDSs lifecycle, has created a felt need for a best practice model for development and evaluation of SLDSs. An obvious first step towards establishing a best practice model is to build a solid overview of current practice. This paper presents a model for the description of SLDS current practice with particular focus on dialogue management and human factors.

The authors present a model of fundamental frequency contour generation, originally developed for speech synthesis applications, by which speakers' intonational behavior can be reduced to a small set of manipulable phonetic parameters. They illustrate the range of interspeaker and intraspeaker variation which this model can produce. Problems with the interpretation of these parameters are also discussed. Emphasis is placed on the use of pitch range to convey speaker characteristics, both relatively short-term emotional states and longer-term personality traits of the speaker.>

This paper describes the prosodic processing and wave-form generation components of the text-to-speech system being developed at Edinburgh University's Centre for Speech Technology Research. Intonation is specified as a sequence of minimal descriptors whose locations are given in terms of syntactically-determined prosodic domains. A pitch contour is computed by converting the descriptors into a sequence of abstract targets whose absolute values depend on a specific speaker model. Duration is determined first at the level of the syllable by a neural network, then accommodated at the segment level according to the distributions observed in a phonetically balanced database. The output waveform is generated by LPC resynthesis of diphone units. Three methods of diphone segmentation are discussed.