5. Overview

Festival is designed as a speech synthesis system for at least three levels of user. First, those who simply want high quality speech from arbitrary text with the minimum of effort. Second, those who are developing language systems and wish to include synthesis output. In this case, a certain amount of customization is desired, such as different voices, specific phrasing, dialog types etc. The third level is in developing and testing new synthesis methods.

This manual is not designed as a tutorial on converting text to speech but for documenting the processes and use of our system. We do not discuss the detailed algorithms involved in converting text to speech or the relative merits of multiple methods, though we will often give references to relevant papers when describing the use of each module.

For more general information about text to speech we recommend Dutoit’s ‘An introduction to Text-to-Speech Synthesis’ dutoit97. For more detailed research issues in TTS see sproat98 or vansanten96.

5.1 Philosophy

One of the biggest problems in the development of speech synthesis, and other areas of speech and language processing systems, is that there are a lot of simple well-known techniques lying around which can help you realise your goal. But in order to improve some part of the whole system it is necessary to have a whole system in which you can test and improve your part. Festival is intended as that whole system in which you may simply work on your small part to improve the whole. Without a system like Festival, before you could even start to test your new module you would need to spend significant effort to build a whole system, or adapt an existing one before you could start working on your improvements.

Festival is specifically designed to allow the addition of new modules, easily and efficiently, so that development need not get bogged down in re-implementing the wheel.

But there is another aspect of Festival which makes it more useful than simply an environment for researching into new synthesis techniques. It is a fully usable text-to-speech system suitable for embedding in other projects that require speech output. The provision of a fully working easy-to-use speech synthesizer in addition to just a testing environment is good for two specific reasons. First, it offers a conduit for our research, in that our experiments can quickly and directly benefit users of our synthesis system. And secondly, in ensuring we have a fully working usable system we can immediately see what problems exist and where our research should be directed rather where our whims take us.

These concepts are not unique to Festival. ATR’s CHATR system (black94) follows very much the same philosophy and Festival benefits from the experiences gained in the development of that system. Festival benefits from various pieces of previous work. As well as CHATR, CSTR’s previous synthesizers, Osprey and the Polyglot projects influenced many design decisions. Also we are influenced by more general programs in considering software engineering issues, especially GNU Octave and Emacs on which the basic script model was based.

Unlike in some other speech and language systems, software engineering is considered very important to the development of Festival. Too often research systems consist of random collections of hacky little scripts and code. No one person can confidently describe the algorithms it performs, as parameters are scattered throughout the system, with tricks and hacks making it impossible to really evaluate why the system is good (or bad). Such systems do not help the advancement of speech technology, except perhaps in pointing at ideas that should be further investigated. If the algorithms and techniques cannot be described externally from the program such that they can reimplemented by others, what is the point of doing the work?

Festival offers a common framework where multiple techniques may be implemented (by the same or different researchers) so that they may be tested more fairly in the same environment.

As a final word, we’d like to make two short statements which both achieve the same end but unfortunately perhaps not for the same reasons:

Good software engineering makes good research easier

But the following seems to be true also

If you spend enough effort on something it can be shown to be better than its competitors.

5.2 Future

Festival is still very much in development. Hopefully this state will continue for a long time. It is never possible to complete software, there are always new things that can make it better. However as time goes on Festival’s core architecture will stabilise and little or no changes will be made. Other aspects of the system will gain greater attention such as waveform synthesis modules, intonation techniques, text type dependent analysers etc.

Festival will improve, so don’t expected it to be the same six months from now.

A number of new modules and enhancements are already under consideration at various stages of implementation. The following is a non-exhaustive list of what we may (or may not) add to Festival over the next six months or so.

• Selection-based synthesis: Moving away from diphone technology to more generalized selection of units for speech database.
• New structure for linguistic content of utterances: Using techniques for Metrical Phonology we are building more structure representations of utterances reflecting there linguistic significance better. This will allow improvements in prosody and unit selection.
• Non-prosodic prosodic control: For language generation systems and custom tasks where the speech to be synthesized is being generated by some program, more information about text structure will probably exist, such as phrasing, contrast, key items etc. We are investigating the relationship of high-level tags to prosodic information through the Sole project http://www.cstr.ed.ac.uk/projects/sole.html
• Dialect independent lexicons: Currently for each new dialect we need a new lexicon, we are currently investigating a form of lexical specification that is dialect independent that allows the core form to be mapped to different dialects. This will make the generation of voices in different dialects much easier.

This document was generated by Alan W Black on December 2, 2014 using texi2html 1.82.