TRAINABLE SPEECH SYNTHESIS
Robert Donovan
June 1996
This thesis is concerned with the synthesis of speech using trainable systems. The research it describes was conducted with two principle aims: to build a hidden Markov model (HMM) based speech synthesis system which could synthesise very high quality speech; and to ensure that all the parameters used by the system were obtained through training. The motivation behind the first of these aims was to determine if the HMM techniques which have been applied so successfully in recent years to the problem of automatic speech recognition could achieve a similar level of success in the field of speech synthesis. The motivation behind the second aim was to construct a system that would be very flexible with respect to changing voices, or even languages.
A synthesis system was developed which used the clustered states of a set of decision-tree state-clustered HMMs as its synthesis units. The synthesis parameters for each clustered state were obtained completely automatically through training on a one hour single-speaker continuous-speech database. During synthesis the required utterance, specified as a string of words of known phonetic pronunciation, was generated as a sequence of these clustered states. Initially, each clustered state was associated with a single linear prediction (LP) vector, and LP synthesis used to generate the sequence of vectors corresponding to the state sequence required. Numerous shortcomings were identified in this system, and these were addressed through improvements to its transcription, clustering, and segmentation capabilities. The LP synthesis scheme was replaced by a TD-PSOLA scheme which synthesised speech by concatenating waveform segments selected to represent each clustered state. The final system produced speech which, though in a monotone, was natural sounding, remarkably fluent, and highly intelligible. The segmental intelligibility was measured using the Modified Rhyme Test, and a 5.0\% error rate obtained. The speech produced by the system mimicked the voice of the speaker used to record the training database. The system could be retrained on a new voice in less than 48 hours, and has been successfully trained on four voices.
N.B. The speech examples referred to in this thesis are available by anonymous ftp to svr-ftp.eng.cam.ac.uk in file pub/reports/donovan_thesis_speech.tar.Z
The tar-file contains a README file and 64 16kHz 16-bit waveform files with 1024 byte NIST headers.
If you have difficulty viewing files that end '.gz',
which are gzip compressed, then you may be able to find
tools to uncompress them at the gzip
web site.
If you have difficulty viewing files that are in PostScript, (ending
'.ps' or '.ps.gz'), then you may be able to
find tools to view them at
the gsview
web site.
We have attempted to provide automatically generated PDF copies of documents for which only PostScript versions have previously been available. These are clearly marked in the database - due to the nature of the automatic conversion process, they are likely to be badly aliased when viewed at default resolution on screen by acroread.