Next Previous Contents

1. Introduction

In the early nineties, the Institute of Phonetics, Faculty of Arts, Charles University, Prague, and the Institute of Radio-Engineering and Electronics, Academy of Sciences, Prague, had managed to assemble a complete TTS implementation for the Czech and Slovak languages, using a linear prediction based diphone synthesis. This TTS engine then served as a base for further research and applications in speech synthesis, until the source became too large and complicated to be easily modified, ported to new hardware or operating system, or to be well understood by anybody except the authors. By the end of 1995, when a need for testing some new prosody modelling hypotheses had arisen, these limitations were slowly becoming a major burden. A new implementation of part of the system was eventually written from scratch (starting in 1996) and it is still expanding, integrating the original results with numerous recent improvements. It has been baptized Epos in 1998.

Our primary design goal is to allow the user the ultimate control over the TTS processing. We avoid hard-wired constants; we use configuration options instead, with sensible default values. Most of the language-dependent processing is driven by a rules file, a text file using an intuitive and well-documented syntax. A rules file lists the rules to be applied on a written text structure representation to yield a corresponding spoken text structure representation (in fact it could be the other way round in principle, but somehow no one seems to need that). Some aspects of user-definable behavior don't fit into the concept of a rules file, and are therefore settable with various options in conventional configuration files. Finally, many other external files can be referenced either by the rules file, or a configuration file, such as segment inventories or dictionaries.

Most of these files have to be processed before any actual TTS processing has finished. That's why Epos is implemented as a background process, i.e. as a daemon under UNIX-like OSes and as a service on Windows NT and similar OSes. Epos reserves a TCP/IP port for any communication with client applications using a custom, quite generic protocol for TTS data flow control, called TTSCP. A simple TTSCP client utility named say-epos has been provided with Epos, but there are many more specialized TTSCP clients in existence.

Epos currently supports several main speech generation algorithms. A linear prediction coding speech synthesizer written by Ellen Víchová allows voice inventories as small as 25 kilobytes, whereas much larger but high-quality voices are available with a time domain synthesizer. Some additional synthesizers not under GPL are used with Epos; you can at least use the virtual speech synthesis to synthesize your own texts using some of them, if you are connected to the Internet. (Your text is partly processed and sent to our server. Then the generated speech signal is sent back to you.) The last option is to use the semi-free MBROLA speech synthesizer, though the MBROLA synthesizer itself is not a part of Epos. We are constantly working on improving the synthesizers.

Epos offers several interesting facilities for prosody generation. The rule-based (yet surprisingly acceptable) prosody based on the prosody research of Zdena Palková offers maximum reliability; the highly flexible neural network framework developed by Jakub Adámek presents a powerful meld of neural network based and rule based tools; and these facilities can also serve as a signal source for a linear prediction based prosody model implemented by Petr Horák. A lot of integration and assessment work in this area has also been done by Daniel Sobe.

The name Epos is not an acronym. It is Greek for...um, go have a look yourself.

This section still has to be (re)written by somebody. At present, try to look at http://epos.ure.cas.cz for additional introductory information or ask the authors by email.

Meanwhile, tell us what kind of introductory information you would like to see here. The documentation is provided for you and we need your feedback.


Next Previous Contents