Abstract:
With the goal of synthesizing natural-sounding speech based on higher-level parameters, variations in subglottal pressure for sentences having different stress patterns have been studied. A Rothenberg mask was used to collect oral pressure and airflow signals from several subjects producing reiterant speech. Acoustic sound pressure was also recorded, and sound pressure levels (SPLs) for both reiterant and natural utterances were compared to verify the suitability of reiterant speech for the task. Subglottal pressure P[inf s] was inferred from oral pressure measured during stop consonant closure of reiterant speech, average glottal flow U[inf g] was estimated from the low-pass-filtered airflow signals, and average glottal area A[inf g] was calculated using P[inf s] and U[inf g]. Glottal source amplitude and spectral tilt were estimated from the speech waveform. Preliminary analysis shows that accented syllables are marked by an increase in P[inf s] of about 50% relative to the weakest unaccented syllables, an increase in SPL of 10 to 15 dB, and some increase in A[inf g]. A model for estimating glottal parameters during these utterances has been developed. Attempts to synthesize sentences based on this model, by including A[inf g] and P[inf s] as control parameters, are described. [Work supported by NIH grants DC00075 and MH52358.]