Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Utility Patent
1999-02-26
2001-01-02
Hudspeth, David R. (Department: 2741)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
C704S251000, C704S009000, C704S010000
Utility Patent
active
06169972
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to an information analysis apparatus and a method for recognizing a user's speech input sentence as a word-lattice in which a plurality of word candidates are linked and for identifying an acceptable word sequences from the word-lattice.
2. Description of the Related Art
Recently, by development of computer technology, a user can input various kinds of information to a computer as a natural method, and a technique to promote the use of the information is required. For example, a computer may be equipped with a microphone and A/D converter. In this case, when the user speaks his request to the computer, the computer recognizes the user's input speech and interpretes his request. In short, the importance of this technique increases.
In the case of speech recognition, as a problem of the prior art, in general, the user's intention in his input speech often does not yield a correct recognition result. In short, in the recognition processing of the input signal, the recognition ratio of “100%” is not expected and a plurality of recognition candidates for one utterance word are often generated. Therefore, when the user utters a sentence, a lattice format in which a plurality of words are linked in order is generated as the recognition result. In order to utilize the user's input by general application program, a post processing technique to quickly determine the correct sentence as the user's input from the recognition result is important.
As the post processing technique, for example, the user's input sentence is determined from the lattice structure of the speech recognition. In this case, as the prior art, a plurality of sentence candidates are generated from the lattice structure and a natural language sentence structure technique is applied for each sentence candidate. For example, in Japanese Patent Disclosure (Kokai) H5-197389, while a lattice of the recognition result obtained from a few word candidates is expanded, an acceptable sentence consisting of simple grammar is identified from the lattice. (In this case, the lattice consists of word candidates and character candidates linked in order from a start node to an end node as the recognition result of the speech input sentence.)
However, in general, the lattice structure includes a large number of word candidates linked in a complicated manner because the input sentence to be recognized includes a large number of vocabulary and complicated grammar used for a dialogue between two persons. For example, for the user's utterance sentence “Nishinomiya-interchange-made-dorekurai-desuka?” [How long does it take to Nishinomiya interchange], a processing result of word-spotting as speech recognition is shown in FIG.
13
. In
FIG. 13
, one word candidate per line is shown as “ID: word candidate (recognition score) word interval in signal frame”. As for the speech input from 0-frame to 40-frame, 42 words including the correct 4 words are obtained. As a result, 38 words are included as false alarms. Assume that the lattice structure obtained from the word spotting result is expanded. Even if a restriction such as the time-appearance position between words is applied, several million word candidates are generated. Therefore, a method to expand the lattice structure is not effective for both calculation quantity and memory quantity.
Furthermore, as an analysis method higher than the expansion method, Japanese Patent disclosure (Kokai) H9-134369 is well known. In this method, an input lattice is expanded as a tree structure in which a head of the lattice is a root, and the tree structure is analyzed by referring to a try-dictionary for quickly analysis. Accordingly, in comparison with the prior method to simply expand all word candidates in the lattice structure, the number of word candidates to be analyzed is reduced because the input lattice is expanded as the tree condition. The analysis is quickly executed by using the try dictionary. However, as for the lattice structure including large ambiguity such as the large number of word candidates as shown in
FIG. 13
, a number of analysis processing iterations greatly increases. In addition to this, the number of retrieval times for the try-dictionary greatly increases while word sequence candidates on interprocessing remaines in the lattice structure.
As mentioned-above, in the prior art, in case of analysis of the lattice structure as a recognition result of the user's speech input sentence, a large quantity of calculation and memory capacity are necessary. As a result, analysis by expanding the input lattice is impossible.
SUMMARY OF THE INVENTION
It is an object of the present invention to provide an information analysis apparatus and method able to quickly extract an acceptable word sequences from the word-lattice as a speech recognition result without expanding the word-lattice.
According to the present invention, there is provided an information analysis apparatus, which connects after a speech recognizer for recognizing a user's speech input sentence and for generating a word-lattice linking a plurality of words from a start node to an end node as recognition candidates of the speech input sentence, for accepting the word-lattice as input and for generating a set of acceptable word sequences by referring to a word-class dictionary that matches each word to corresponding word-class and to a grammar that matches each word-class sequence to corresponding sentence type, comprising: a hash dictionary means for storing a plurality of identifiers of word-class sequences, each of which represents a sentence by unit of the word-class and word appearance number, the word-class being positioned at the word appearance number in the word-class sequence; an initialization means for forming a node for each word in the word-lattice, the node consisting of the word-class, an interprocessing list, lists of next nodes and unprocessed antecedent nodes in the word-lattice, the interprocessing list of the node directly linked from the start node representing the identifiers of word-class sequences for corresponding word-class and the word appearance number “1” in said hash dictionary means, and for forming a list of processing nodes representing the nodes directly linked from the start node; a propagation means for extracting one node from the list of processing nodes if the list of unprocessed antecedent nodes of the one node is empty, for extracting each next node of the one node from the list of next nodes if the list of next nodes is not empty, for retrieving the identifiers of word-class sequences from said hash dictionary means by the word appearance number as link order and the word-class of the each next node, for respectively calculating a product of retrieved identifiers of the each next node and the identifiers in the interprocessing list of the one node, for storing the product as propagated identifiers in the interprocessing list of the each next node, for deleting the one node from the list of unprocessed antecedent nodes of the each next node and from the list of processing nodes, and for adding the each next node in the list of processing nodes; a repeat means for repeating the process of said propagation means until the list of processing nodes is empty; and word sequence extraction means for extracting the propagated identifiers of the end node if the list of processing nodes is empty, and for extracting the word sequences corresponding to the word-class sequences of the propagated identifiers from the word-lattice.
Further in accordance with the present invention, there is provided an information analysis apparatus, which connects after a speech recognizer for recognizing a user's speech input sentence and for generating a word-lattice linking a plurality of words from a start node to an end node as recognition candidates of the speech input sentence, for accepting the word-lattice as input and for generating a set of acceptable word
Kono Yasuyuki
Yano Takehide
Azad Abul K.
Finnegan Henderson Farabow Garrett & Dunner L.L.P.
Hudspeth David R.
Kabushiki Kaisha Toshiba
LandOfFree
Information analysis and method does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Information analysis and method, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Information analysis and method will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2469583