Publications by Matthew Aylett

This paper first introduces a newly-recorded high quality Romanian speech corpus designed for speech synthesis, called “RSS”, along with Romanian front-end text processing modules and HMM-based synthetic voices built from the corpus. All of these are now freely available for academic use in order to promote Romanian speech technology research. The RSS corpus comprises 3500 training sentences and 500 test sentences uttered by a female speaker and was recorded using multiple microphones at 96 kHz sampling frequency in a hemianechoic chamber. The details of the new Romanian text processor we have developed are also given. Using the database, we then revisit some basic configuration choices of speech synthesis, such as waveform sampling frequency and auditory frequency warping scale, with the aim of improving speaker similarity, which is an acknowledged weakness of current HMM-based speech synthesisers. As we demonstrate using perceptual tests, these configuration choices can make substantial differences to the quality of the synthetic speech. Contrary to common practice in automatic speech recognition, higher waveform sampling frequencies can offer enhanced feature extraction and improved speaker similarity for HMM-based speech synthesis.

Keywords: Speech synthesis, HTS, Romanian, HMMs, Sampling frequency, Auditory scale

Unit selection speech synthesis has reached high levels of naturalness and intelligibility for neutral read aloud speech. However, synthetic speech generated using neutral read aloud data lacks all the attitude, intention and spontaneity associated with everyday conversations. Unit selection is heavily data dependent and thus in order to simulate human conversational speech, or create synthetic voices for believable virtual characters, we need to utilise speech data with examples of how people talk rather than how people read. In this paper we included carefully selected utterances from spontaneous conversational speech in a unit selection voice. Using this voice and by automatically predicting type and placement of lexical fillers and filled pauses we can synthesise utterances with conversational characteristics. A perceptual listening test showed that it is possible to make synthetic speech sound more conversational without degrading naturalness.

In speech synthesis the unit inventory is decided using phonological and phonetic expertise. This process is resource intensive and potentially sub-optimal. In this paper we investigate how acoustic clustering, together with lexicon constraints, can be used to build a self-organised inventory. Six English speech synthesis systems were built using two frameworks, unit selection and parametric HTS for three inventory conditions: 1) a traditional phone set, 2) a system using orthographic units, and 3) a self-organised inventory. A listening test showed a strong preference for the classic system, and for the orthographic system over the self-organised system. Results also varied by letter to sound complexity and database coverage. This suggests the self-organised approach failed to generalise pronunciation as well as introducing noise above and beyond that caused by orthographic sound mismatch.

The ability to use the recorded audio of a subject’¡Çs voice to produce an open-domain synthesis system has generated much interest both in academic research and in commercial speech technology. The ability to produce synthetic versions of a subjects voice has potential commercial applications, such as virtual celebrity actors, or potential clinical applications, such as offering a synthetic replacement voice in the case of a laryngectomy. Recent developments in HMM-based speech synthesis have shown it is possible to produce synthetic voices from quite small amounts of speech data. However, mimicking the depth and variation of a speaker’¡Çs prosody as well as synthesising natural voice quality is still a challenging research problem. In contrast, unit-selection systems have shown it is possible to strongly retain the character of the voice but only with sufficient original source material. Often this runs into hours and may require significant manual checking and labelling. In this paper we will present two state of the art systems, an HMM based system HTS-2007, developed by CSTR and Nagoya Institute Technology, and a commercial unit-selection system CereVoice, developed by Cereproc. Both systems have been used to mimic the voice of George W. Bush (43rd president of the United States) using freely available audio from the web. In addition we will present a hybrid system which combines both technologies. We demonstrate examples of synthetic voices created from 10, 40 and 210 minutes of randomly selected speech. We will then discuss the underlying problems associated with voice cloning using found audio, and the scalability of our solution.

In speech synthesis the inventory of units is decided by inspection and on the basis of phonological and phonetic expertise. The ephone (or emergent phone) project at CSTR is investigating how self organisation techniques can be applied to build an inventory based on collected acoustic data together with the constraints of a synthesis lexicon. In this paper we will describe a prototype inventory creation method using dynamic time warping (DTW) for acoustic clustering and a joint multigram approach for relating a series of symbols that represent the speech to these emerged units. We initially examined two symbol sets: 1) A baseline of standard phones 2) Orthographic symbols. The success of the approach is evaluated by comparing word boundaries generated by the emergent phones against those created using state-of-the-art HMM segmentation. Initial results suggest the DTW segmentation can match word boundaries with a root mean square error (RMSE) of 35ms. Results from mapping units onto phones resulted in a higher RMSE of 103ms. This error was increased when multiple multigram types were added and when the default unit clustering was altered from 40 (our baseline) to 10. Results for orthographic matching had a higher RMSE of 125ms. To conclude we discuss future work that we believe can reduce this error rate to a level sufficient for the techniques to be applied to a unit selection synthesis system.

In commercial systems the memory footprint of unit selection systems is often a key issue. This is especially true for PDAs and other embedded devices. In this year's Blizzard entry CereProc R gave itself the criteria that the full database system entered would have a smaller memory footprint than either of the two smaller database entries. This was accomplished by applying Speex speech compression to the full database entry. In turn a set of small database techniques used to improve the quality of small database systems in last years entry were extended. Finally, for all systems, two quality control methods were applied to the underlying database to improve the lexicon and transcription match to the underlying data. Results suggest that mild audio quality artifacts introduced by lossy compression have almost as much impact on MOS perceived quality as concatenation errors introduced by sparse data in the smaller systems with bulked diphones.