Formant diphone parameter extraction utilising a labelled single-speaker database

Abstract

This paper examines a method for formant parameter extraction from a labeled single speaker database for use in a formant-parameter diphone-concatenation speech synthesis system. This procedure commences with an initial formant analysis of the labelled database, which is then used to obtain formant (F1-F5) probability spaces for each phoneme. These probability spaces guide a more careful speaker- specific extraction of formant frequencies. An analysis-by-synthesis procedure is then used to provide best-matching formant intensity and bandwidth parameters. The great majority of the parameters so extracted produce speech which is highly intelligible and which has a voice quality close to the original speaker. Synthesis techniques based upon LPC-parameter or waveform concatenation are much less vulnerable to the effects of poorly extracted parameters. The formant model is, however, more straightforwardly related to the source-filter model and thus to speech production. Whilst it is true that overlap-add concatenation of waveform-based diphones can easily model a voice with quite high fidelity, new voices and voice qualities require the recording of new speakers (or the same speaker utilising a different voice quality) and the extraction of a new diphone database. Such systems can be used to examine the effects of intonation and rhythm on voice quality or vocal affect but formant-based systems can much more readily examine the effect of frequency-domain modifications on voice quality. Such modifications might include formant frequency shifting, bandwidth modification, modification of relative formant intensities and spectral slope variation. It is even possible, if the synthesiser design allows it, to experiment with the insertion of additional poles and zeroes into the spectrum such as might occur when modelling the "singer's formant" for certain styles of singing voice. Such research requires a parallel formant synthesiser with a great deal of flexibility of control. Further, and most importantly, it requires a diphone database that is extremely accurate. Formant errors must be minor and few in number and this should be achieved without excessive hand correction. Formant tracks should display, as far as possible, pole continuity across fricatives, stops and affricates. Extracted intensities and