Automatic phonetic baseform determination

Low-cost implementation of open set keyword spotting

“…Text-based systems produce more consistent pronounciations than spoken input systems, but have difficulty modelling proper nouns. The resulting combined strings are far more robust, and previous work has shown that recognition performance equal to the level achieved using hand-transcribed strings is possible (Bahl et al, 1991).…”

“…Ideally the user should be able to just state the keyword once. The simplest representation of the KPS is the optimal Viterbi path recognized (Bahl et al, 1991;Asadi et al, 1991;Wilcox & Bush, 1991). This may not produce a good match of the keyword due to recognition errors and the fact that a single utterance does not necessarily represent the general pronunciation of the keyword.…”

“…In practice, it is expected that the user might be prepared to provide a maximum of three samples. To produce the KPS the multiple-candidate transcription method (Bahl et al, 1991) is used. A set of possible transcriptions is obtained by performing n-best recognition on each keyword utterance.…”

“…Secondly, many word-spotting applications are user specific, such as personal memo and dictation machines. Retrieval requests are, therefore, likely to contain a high occurrence of proper nouns and user-specific jargon, both of which tend to have highly irregular pronunciations (Bahl et al, 1991). Thus, it is difficult to get complete coverage of such terms in spelling based systems, which rely on either a pronunciation dictionary or a text-to-phone mapper (see Klatt, 1987;Damper, 1995;Meng, 1995, for reviews of the literature).…”