The Speect system can be divided into two distinct parts, the engine and plug-ins. The Speect engine is completely independent of any text-to-speech specific modules, and is solely responsible for the loading of voices and their associated data and plug-ins, and controlling the synthesis process. The plug-ins provide new object types and interfaces to voice data (linguistic and acoustic), as well as the processors that convert some form of textual input into a waveform.
Speect is not meant to replace speech processing toolkits. The linguistic and acoustic models and data still needs to be generated by packages such as Edinburgh Speech Tools , Festvox , the Speech Signal Processing Toolkit  and others.
For this introduction we give a very brief overview of the main functional blocks of the text-to-speech process, and a high level description of how these abstract functional blocks are implemented in Speect.
Text-to-speech (TTS) synthesis is the automated process of mapping a textual representation of an utterance into a sequence of numbers representing the samples of synthesized speech . This conversion is achieved in two stages as depicted in figure 1:
- Natural Language Processing (NLP): Converting the textual representation of an utterance into symbolic linguistic units.
- Digital Signal Processing (DSP): Mapping the symbolic linguistic units into samples of synthesized speech.
The Natural Language Processing stage consists of the following major modules:
- Text pre-processing involves the transformation of the textual input into a format suitable for the phonetization module. The specifics of this task is dependent on the type of textual input given to the system and includes utterance chunking and text normalization.
- The normalized text of the pre-processing module is converted into a phonetic representation by the phonetization block.
- Prosody generation involves the generation of intonation and duration targets through some form of prosody models.
The data generated by the NLP stage represents the symbolic linguistic units, which are then converted into synthetic speech by the Digital Signal Processing stage. The DSP stage can be realized by means of unit selection , statistical parametric synthesis , formant synthesis , or some other type of synthesizer technology. Each of the modules in the two stages adds some type of information to the initial given utterance which enables the final module, waveform generation, to generate synthetic speech based on this information. The NLP stage is language dependent, whereas the DSP stage is dependent on the synthesizer technology of the implemented synthetic voice.
From the functional blocks or modules of the NLP and DSP stages of figure 1 we define the following objects and processes:
The utterance is the input and output of all the functional blocks shown in figure 1, even for the DSP processors which generate the speech signal. Speect models the utterance internally as a heterogeneous relation graph , and all the modules of figure 1 just add information to the utterance.
In Speect the function blocks depicted in figure 1, which do some or other conversion to their respective inputs, are known as utterance processors. An utterance processor receives an utterance as input and transforms the utterance in some or other way based on knowledge of the:
- input type: an email message, for example, requires some extra processing when compare to a single line of text,
- language: phonetization will for example be different for English, when compared to isiZulu, and the
- voice: different voices will have different speaking rates, pitch contours and so on.
From figure 1, we can see that there is a pipeline of utterance processors doing transformations on the utterance, and producing the synthetic speech. We call this collection of utterance processors an utterance type.
Utterance types can be defined for any input types, languages or synthesizer technologies, by just having a different pipeline of utterance processors.
Utterance processors also make use of feature processors. A feature processor extracts features from individual units in an utterance. Feature processors are defined in a key-value (name - processor implementation) mapping, and are called by their names. The real power of feature processors becomes apparent when doing multilingual TTS, for example, we can reuse utterance processors and just redefine the key-value pair of a feature processor (same name, different implementation).
All of this comes together in the definition of a voice. In Speect, a voice defines the utterance types that can be used for synthesis with the specific voice. Each of these utterance types defines a pipeline of utterance processors. The voice also defines the feature processors key-value mapping, connecting a named feature processor to a specific implementation, which the utterance processors then can use. Finally, the voice defines it’s data, be that linguistic (phonesets, grapheme to phoneme rules, ...) or acoustic (unit inventory, acoustic models, ...). Figure 2 shows a representation of this voice definition.
|||The Centre for Speech Technology Research, The University of Edinburgh, The Edinburgh Speech Tools Library, http://www.cstr.ed.ac.uk/projects/speech_tools/.|
|||Black, A. and Lenzo, K. “Building Voices in the Festival Speech Synthesis System”, http://www.festvox.org/festvox/bsv.ps.gz, 2003.|
|||Department of Computer Science, Nagoya Institute of Technology, Speech Signal Processing Toolkit, http://sp-tk.sourceforge.net/.|
|||Stylianou, Y., “Harmonic plus noise models for speech,combined with statistical methods, for speech and speaker modification”, Ph.D. Thesis, Ecole Nationale Superieuredes Telecommunications, Paris, France, 1996.|
|||Hunt, A. and Black, A. “Unit selection in a concatenative speech synthesis system using a large speech database”, In Proceedings of ICASSP, vol 1, pp. 373-376, Atlanta, Georgia, 1996.|
|||Black, A., Zen, H., and Tokuda, K. “Statistical Parametric Synthesis”, Proceedings of ICASSP, Hawaii, 2007.|
|||Högberg, J. “Data driven formant synthesis”, In Proceedings of Eurospeech, pp. 565-568, Greece, 1997.|
|||Taylor, P., Black, A.W., and Caley, R. “Heterogeneous relation graphs as a mechanism for representing linguistic information”, Speech Communication 33:153-174, 2001.|