Text-to-Speech Architectures

One of the earliest and most common encodings was ASCII (the American Standard Code for Information Interchange), which gave a character value to the first 127 values in an 8-bit byte of memory (the final bit was left unused). Ascii is of course a standard that was developed in the United States, which could only represent the 26 characters of the standard English alphabet. Most of the worlds languages of course require difierent characters and so ascii alone will not suffice to encode these. A series of extensions were developed primarily with use with other European characters, often by making use of the undefined eighth bit in ascii. 

Unicode simply defines a number for each character, it is not an encoding scheme in itself For this, a number of different schemes have been proposed. UTF-8 is popular on Unix machines and on the internet. Is is a variably width encoding, meaning that normal ascii remains unchanged but that wider character formats are used when necessary. By contrast UTF-16 popular in Microsoft products, uses a fixed size 2 byte format. More recent extensions to Unicode mean that the original 16 bit limitation has been surpassed, but this in itself is not a problem (specifically for encodings such as UTF-8 which are extensible). 

The various algorithms described here all produce output, and this has to be passed to the next module in the TTS system. As the next module will produce output too, it makes sense to have a general mechanism for storing and passing information in the system. This brings us to the issue of how to design our text-to-speech architecture. 

Most TTS systems have adopted a solution whereby a single data structure is passed between each module. Usually, this data structure represents a single sentence, so the TTS system works by first splitting the input into sentences by the sentence splitter, forming a data structure containing 

In many early systems, the utterance structure was no more than a string that was passed between modules. In doing this systems could adopt either an overwrite paradigm or a addition paradigm. For instance in an overwrite system if the input text was 

The man lived in Oak St. the text-normalisation module might produce ouQ)ut like 

The point is that, in each case, the module takes its input and ontputs only file particular results from that modifie. In an addition paradigm, each module adds its output to the output from the previous modules. So in the above example, the outpnt from the text-normalisation modnle might produce something like 

---