5. Established Methods

Both pre-processing (feature extraction) and pattern classification techniques are discussed in this chapter. Traditionally, specialised parameters have been used for the analysis of speech disorders: harmonic-to-noise ratio, jitter, shimmer, and others. These have been devised using expert opinions from speech and language therapists and other professionals. They are typically calculated using widely available software packages, but still require trained personnel to collect and prepare the recordings, as well as to interpret the resulting parameters. More recently, researchers have also investigated many of the parameters or features used in speech and speaker recognition. Features such as the ubiquitous mel-frequency cepstral coefficients are often used, but so are numerous less common methods, such as formant frequencies, modulation spectra, chaos-theory parameters, and prosodic and phonological features. Each of these has had its fair share of success, but the most successful systems have generally used a combination of multiple features and/or multiple classification algorithms. Numerous methods for discriminating between disordered and normal speech, and sometimes between different forms of speech disorder, have been devised. They have typically been based on neural networks, Markov models, support vector machines, and other classifiers (both linear and non-linear), although Gaussian Mixture Models are probably the most widely used, robust, and successful so far.