JPH09244688A - Speech recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To carry out highly precise recognition utilizing environment-dependent phoneme standard patterns. SOLUTION: Standard patterns are determined while a network is generated by making good use of context free grammar having an environment-dependent phoneme as a termination symbol and an LR purser, and a one-pass search is advanced while an input speech is matched against standard patterns of HMM; and environment-dependent phoneme standard patterns are used as the standard patterns of HMM, a phoneme of each arc is propagated to a node which is precedent by the number of the said environment-dependent phonemes at the time of the extension of the network, and at the time of arc extension, the phoneme standard pattern of a triphone model is determined at a tip node 3 with a phoneme P3 predicted by the LR purser and phoneme series P1 and P2 having reached the node 3 and set to the arc between nodes 3 and 4.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a hidden Markov model (for example, Seiichi Nakagawa “Speech Recognition by Probabilistic Model” edited by The Institute of Electronics, Information and Communication Engineers (1988)) and a generalized LR context analysis algorithm (for example, M. Tomita: " An Efficient Context-F
ree Parsing Algorithm for Natural Languages and it
s Applications ”, Carnegie-Mellon University (198
5)) and the One-Pass search algorithm (eg J. Br.
idle and others: "An Algorithm For Connected Word Recogni
, ICASSP82, pp. 899-902 (1982)) and a method for continuously recognizing input speech.

[0002]

2. Description of the Related Art In conventional speech recognition, an LR parser (grammar analyzer) is used to search for a recognition candidate that meets a predetermined grammar, and recognition is performed while matching a standard pattern of the recognition candidate with an input speech. There is a way to do. Also, using a predetermined network, a part of the network is compared with the standard pattern and the input voice, and except for those with a small score, the search is advanced while minimizing the increase in the processing amount, and finally the score A One-Pass search method in which the highest candidate is used as the recognition result is known.

Furthermore, a method of speech recognition is known in which a hidden Markov model, a generalized LR parser, and a One-Pass algorithm are integrated. As this integrated recognition method, a sub-network is created from the context-free grammar with environment-independent phonemes using the LR parser, and the environment-dependent phonemes are considered at the junctions between these sub-networks to determine the environment-dependent phoneme standard. From methods using patterns (for example, Katsutoshi Ito: “Continuous Speech Recognition Using Extended LR Parsing Method”, IEICE Technical Report, SP90-74, pp. 49-56) and context-free grammars using environment-independent phonemes. There was a method to dynamically create one network using a parser and use the phoneme standard pattern independent of environment. (For example, Kenji Kita et al .: "L
One-Pass continuous speech recognition algorithm by R parser control ", IEICE Tech., SP94-13 (1994-
05) 55-62).

From the viewpoint of efficiency, it is better to process with one network without using the sub-network, and from the viewpoint of accuracy of speech recognition, it is better to use the environment-dependent phoneme standard pattern, but both are used at the same time. There was no way I could do it.

[0005]

SUMMARY OF THE INVENTION The present invention integrates a Hidden Markov Model, a generalized LR parser, and a One-Pass algorithm, efficiently processes them using a single network, and at the same time, uses environment-dependent phoneme standard patterns. It is an object of the present invention to provide a voice recognition method that can utilize the above and has a high recognition rate.

[0006]

According to the present invention, by using the environment-dependent phoneme standard pattern, the phonemes of each arc of the network are transmitted to each node up to the preceding node by at least the number taking the phoneme environment dependence into consideration. For the newly propagated node on the network, the phoneme predicted by the LR parser to be set to the arc to that node and the past phoneme sequence stored at the source node of that arc , The environment-dependent phoneme standard pattern to be set for the arc is determined.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 functionally shows the structure of a recognition device to which an embodiment of the present invention is applied. The voice input from the input terminal 1 is converted into a digital signal in the feature extraction unit 2 and further subjected to LPC cepstrum analysis, and then 1
It is converted into a feature parameter every frame (for example, 10 milliseconds). This feature parameter is, for example, an LPC cepstrum coefficient.

An environment-dependent phoneme standard pattern of a hidden Markov model is created from the learning speech database in the same format as the above-mentioned feature vector and stored in the standard pattern memory 4. From the context-free grammar that defines the range of the recognition target and uses the environment-independent (environment-independent) phonemes as terminal symbols,
A table for the R parser is created and stored in the LR table 5.

FIG. 2 shows a flowchart of the processing in the recognition section 3.
Shown in The characteristic parameters of the input speech for one frame are read (S ₁ ), the network is grown (every frame or every several frames) (S ₂ ), and pattern matching between the input speech characteristic parameters and the standard pattern is performed (S 1).
₃ ) Then, in the next step, the active node to be processed is determined and the process returns to step S ₁ (S ₄ ). This is repeated until there is no input voice. Even if the input voice does not disappear, the process may be terminated under some condition from the score obtained by the pattern matching.

The network consists of nodes and arcs, initially there is only one initial node. Each node is a hypothesis generated by using the LR parser, and one having the same stack contents is assigned and associated. Therefore, LR parsers that perform the same processing thereafter are integrated into the same node (arcs are connected to the same node). Here, this is called a merge node. Based on the collation result in each step, the one with a high score is set as an active node to be processed in the next step. Other nodes may be retained in the network as they are, or may be discarded for effective use of memory. In this way, the One-P is extended while expanding the network by using the context-free grammar and the LR parser, which are environment-independent phoneme termination signals.
We will proceed with the search by the ass search.

The generation (growth operation) of the network, which is a characteristic part of the present invention, will be described with reference to FIG. Since a network of infinite size can be created from the context-free grammar, it is assumed that the entire network cannot be created in advance. So in the network generator,
The operation of creating a new node from each active node of the network using the LR parser and extending the arc (growing the network) is limited to a finite number of phonemes ahead. Here, a case where one phoneme is extended at a time will be described. In the figure, the symbols shown above the dotted lines are environment-independent phonemes predicted between the nodes by the LR parser. A solid line between the nodes is an arc of the network. In the present invention, an environment-dependent phoneme standard pattern is set (1-2, 2-3, in order to make the diagram easy to see).
Arcs between nodes 5-6 and 6-3 are omitted). Here, the so-called triphon, which takes into consideration the environment on the left side (L) and the right side (R) of the central phoneme (C).
This will be described in the case of the e model and will be referred to as LCR. When a node is set, the phonemes set in the arc between the previous nodes are propagated to the new node. In this example, trip
Since this is the case of the phone model, the phonemes of each arc to the preceding node are propagated and held.

First, a simple case of FIG. 3A will be described. No
Nodes 1 to 3 from the generated network
Then, consider the case of growing the network. LR Par
The phoneme P that is environment independent from node 3_ThreeIs predicted
And change to the state of the stack represented by the new node 4.
Suppose you know that Where it is propagated to node 3
Information of the past two phoneme series stored as (P₁, P_Two)
From P₁, P_Two, P _ThreeA three phoneme sequence is obtained.
Based on this, the arc between nodes 3-4 is₁−P _Two−P_Three
The environment-dependent phoneme standard pattern is set.

Diagram of the case where node 3 is a merge node
3B will be described. In this case, it is propagated to node 3 and stored.
As the obtained past two-phoneme series, (P₁, P_Two) And (P
_Four, P_Five) Exists. New by LR parser
Predicted environment-independent phoneme P _ThreeIn combination with
P between nodes 3-4₁−P_Two−P_ThreeAnd P_Four−P_Five−
P_ThreeIt turns out that we need two arcs to set. So
Then, in order to ensure acoustic compatibility, two node 3
Or P for the arc from node 2₁−P _Two−P_Three
Is P for the arc from node 6_Four−P_Five−P_ThreeSupported by
Restrictions are set so that the data will be passed.

In order to reduce the processing amount, it is possible to consider a method that does not consider consistency. In this case, the result obtained contains an error. Past phoneme series up to node 3 and node 3
-4 instead of being referred to as a new phoneme combination, the entire phoneme sequence required for node 4, P ₁ − in the above example
P ₂ −P ₃ (and P ₄ −P ₅ −P ₃ ) may be held and referred to, or may be propagated and held from the first node including an unnecessary phoneme sequence, and only a necessary portion may be held. You may refer.

Finally, on the LR parser,
The pt) hypothesis becomes one merge node. At the end of all steps, the recognition result can be obtained by tracing back the arc that propagated the highest score to this merge node. The environment dependence of phonemes is not limited to the triphone model, and the number of phonemes to be propagated and held by each node is
The number is at least one less than the number of phonemes in the environment-dependent standard pattern.

[0016]

EFFECT OF THE INVENTION Suppose that an environment-dependent phoneme is a terminal symbol.
When the R table is used, the number of merge nodes is small, and although the syntax is the same, the network grows while being divided into a plurality of paths, resulting in a large number of paths, that is, a large number of candidates and a large amount of calculation. However, in the present invention, the merge node which is distributed when the LR table having the environment-dependent phoneme as the terminal symbol is dispersed is avoided by using the LR table having the environment-independent (environment-independent) phoneme as the terminal symbol. Then, the environment-dependent phoneme standard pattern is determined using the phonemes propagated by the merge node as many times as necessary, and it is set in the arc of the newly created node. It is possible to perform high-accuracy recognition processing using environment-dependent phoneme standard patterns while dynamically creating.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a flowchart of a recognition processing unit showing an embodiment of the present invention.

FIG. 3 is a diagram illustrating a method of setting an environment-dependent phoneme standard pattern in an arc of a network according to an embodiment of the present invention.

Claims (1)

[Claims] 1. A phoneme standard pattern is determined while generating a network by using an LR parser and a context-free grammar whose input speech is a characteristic parameter time series and environment-independent phonemes are terminal symbols. One-while checking the characteristic parameter time series and the standard pattern of the hidden Markov model
In a speech recognition method in which a search is advanced by a Pass search and candidates having a high similarity likelihood are used as recognition results, an environment-dependent phoneme standard pattern is used as a standard pattern of the hidden Markov model, and at least phonemes of each arc of the network are used. Propagate to the nodes up to the previous node by the number considering the phoneme environment dependence, and propagate to the phoneme newly predicted by the LR parser and the arc reaching that node at the tip node when expanding the above network. A speech recognition method characterized in that an environment-dependent phoneme standard pattern is determined based on a phoneme sequence that has been obtained and is set in an arc on a network.