The use of phonetic labelling
In the sections on Acoustic Phonetics and Symbols and Spelling, you have been told how symbols can be attached to acoustic waveforms and saved as annotated files. In doing practical work, you will have found that this is difficult and time-consuming. Yet many phonetics experts have spent enormous amounts of time doing this. What is the reason for doing this work?
The answer lies in the statistical study of speech. One of the biggest problems in the scientific study of speech is that of variability. Everyone speaks differently from everyone else, and nobody says something exactly the same way twice. Therefore if we wish to find out the essential nature of the sounds of a language or dialect we need to have a systematically collected corpus or database of recordings of that language, together with phonetic labelling. In this way, we can find, for example, 1000 examples of the /i/ vowel or the /g/ consonant of a language and use a computer to figure out what its essential acoustic characteristics are.
This becomes particularly important in the field of Automatic Speech Recognition. We need to be able to present the computer with, for example, 10,000 examples of the /e/ phoneme in different contexts and spoken by different speakers so that the computer can learn to recognize that particular sound as effectively as possible.
A good example of a collection of speech recordings as described above was produced by a number of members of the present PHON2 group. This research project, which was given the name BABEL, was funded by the European Union and produced a labelled database of Bulgarian, Estonian, Hungarian, Polish and Romanian. There is a web-site where you can read about this project: click here. The database can be bought from ELRA (an organization run by the EU which sells material of this sort): click here.
There are several types of program which learn to recognize speech sounds by using labelled speech databases. One is the technique of Hidden Markov Models. The other is that of Neural Networks. Studying these techniques is difficult unless you are expert in mathematics, but you can read some material about them: click here.