| 摘要 |
<p>The method involves sampling a test speech signal at regular time points, deriving test signals from the sampled values, comparing with reference values to derive evaluation values for detecting words, whereby sequences of reference signals represent words and are derived from a a training speech signal. Speech models derived from the training results give the probabilities for sequences of words and are used to detect words. A new vocabulary is generated from the training speech signal which contains set elements of words and word groups which occur in the training speech signal with at least a defined a frequency. New speech models are derived for the vocabulary from the training speech signal and are used with the new vocabulary to detect words from the test speech signal.</p> |
