AI Case Study

"CereProc has developed a core speech synthesis engine, CereVoice.

CereProc's unit selection voices are built from large databases of recorded speech, during database creation, each recorded utterance is segmented into some or all of the following: individual phones, syllables, morphemes, words, phrases, and sentences. The division into segments is done using a specially modified speech recognizer.[1] An index of the units in the speech database is then created based on the segmentation and acoustic parameters like the fundamental frequency (pitch), duration, position in the syllable, and neighbouring phones. At runtime, the desired target utterance is created by determining the best chain of candidate units from the database (unit selection). Unit selection provides the greatest naturalness, because it applies digital signal processing (DSP) to the recorded speech only at concatenation points. DSP often makes recorded speech sound less natural.

CereProc's parametric voices produce speech synthesis based on statistical modelling methodologies; in this system, the frequency spectrum (vocal tract), fundamental frequency (vocal source), and duration (prosody) of speech are modelled simultaneously. Speech waveforms are generated from these parameters using a vocoder. Critically, these voices can be built from significantly less recorded speech than unit selection voices and have a much smaller footprint when installed, because of this they are used for private voice cloning."