6.2
Speech analysis and classification
The analysis of speech is an important requirement of many different applications and the
classification of speech into various categories is a necessary part of many techniques.
A full list would be lengthy, but the following subset of basic techniques indicates the
sheer range of applications of speech analysis and classification:
• detecting the presence of speech;
• detecting voiced or unvoiced speech;
• finding boundaries between phonemes or words;
• classifying speech by phoneme type;
• language detection;
• speaker recognition;
• speech recognition.
Classification is an important, and growing area of speech research which relates to
the machine ‘understanding’ of speech (where understanding can range from know-
ing whether speech is present right through to understanding the meaning or emotion
conveyed by the spoken word).
In order to begin such classification of speech, it is usually necessary to first perform
some form of measurement on the speech signal itself. For example detecting voiced
or unvoiced speech might require the determination of speech power and pitch, perhaps
through examination of LSP data. Many methods can potentially be used for analysis of
speech, and extensive empirical testing is almost always required to determine the best
subset of measures to be used for a particular application, whatever that may be.
By and large, new classification tasks will make use of many of the same basic speech
measures as described in our analysis toolkit (Section 6.1). Measures can be used in
different ways, to different ends, and most often several measures must be combined
together.
In addition to the several methods in our toolkit, there are some methods which
have been very much reserved for speech analysis, particularly pitch detection-related
methods, since pitch is so important to speech communications.
In the following subsections, we will examine techniques for the analysis and extrac-
tion of pitch from speech.
Do'stlaringiz bilan baham: |