word boundary- hypothesisation in hindi speech - Speech and ...

syllables than at word boundaries preceding strong syllables, indicating that the
speakers are aware of the use of strong syllables to indicate word beginnings. Thus
these findings confirmed the use of strong syllables to detect word boundaries by
humans and validate the use of MSS.
The above studies established that prosodic knowledge can be used to
hypothesise word boundaries in speech. The studies have also identified some prosodic
features, such as pause, duration and pitch(FO), as possible clues to word boundaries.
Our studies for Hindi using these prosodic features for word boundary hypothesisation
are reported in chapter 6.
2.3.4 Word boundary hypotheskation techniques based on acoustic-phonetic knowledge
Not many word boundary detection techniques were reported in literature in
which the acoustic-phonetic knowledge was explicitly used. However, two techniques
were reported which used spectral information to detect the word boundaries. Both
operate directly on the speech signal and hypothesise word boundaries in it.
The first technique was developed for application in a connected word
recognition task [Zelenski and Class 19831. It used an algorithm which was based on
estimation principles. In this the input speech signal was divided into a sequence of
windows. The signal in the window was represented by a parameter vector x =
{x1,x2, ... xL), where each of the xi represent a speech parameter such as one of the
outputs of a filter bank. The word boundary hypothesisation problem was posed as one
of classifying a given window into one of the two classes: (i) class1 window, containing
a word boundary, and (ii) class2 window, not containing a word boundary. Ideally the
classifier should produce an output z, where z = 1 for window class 1, and z = 0 for
window class 2.
The target value z can be approximated by an estimation d which is computed