Method and system for recognizing speech of agglutinative language

Abstract

The embodiment of the invention relates to a method and a system for recognizing Korean speech. The method comprises the following steps of extracting speech long-term features; calculating the posterior probability of an extension phoneme set for the long-term features; performing PCA (Principal Component Analysis) dimension reduction processing on the posterior probability to acquire MLP (Multilayer Perception) features based on the long-term features; and performing speech recognition based on a gaussian mixture model-hidden markov model (GMM-HMM) framework on the MLP features, and thus acquiring a recognition result. According to the method and the system, the Korean phoneme set is detailed and classified by means of the advantages of the long-term features in the aspect of depicting collaborative pronunciation, the confusion degree of acoustic models is effectively reduced, and an effect of recognizing the speech is improved.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to a method and system for agglutinative speech recognition. Background technique [0002] Adhesive language mainly relies on the change of word endings to express its grammatical relationship, and its typical feature is that word-level units in the language are composed of a large number of morpheme connections. Compared with Chinese, which belongs to the analytical language, the sticky feature brings many new challenges to speech recognition, which greatly affects its performance under the traditional speech recognition framework. Among these new challenges, the more important ones can be summarized as the following two points: The first is that in terms of language model modeling, Korean natural language units such as words and words separated by spaces are not suitable for language model modeling unit; the second is in terms of acoustic model modeling, the severe co-pronunciat...

AI Technical Summary

Problems solved by technology

There is still no effective solution to the second challenge. Some researchers have tried to reduce the degree of confusion of the acoustic model by introducing the concept of isotopes under short-term characteristics, but experiments have proved that this method is not effective on the basic single-factor modeling unit. The effect is obvious, but the effect is not ideal on the three-factor modeling unit used in conventional speech recognition systems

Images