JP2006171096A - Continuous input speech recognition device and continuous input speech recognizing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To make it unnecessary to perform speech segmentation in a speech input stage with respect to continuous input speech becoming objects of recognition, and also to improve recognition rate. <P>SOLUTION: A continuous input speech recognition device of the present invention is a device which performs recognition of input speech using a sound model in which features of phonemes are held, and a language model in which contents of utterance are held and a dictionary in which correspondence among phonemes and morpheme elements are held, and the device is equipped with; the language model in which a plurality of sentences are connected with pause signs; the sound model in which features of phonemes each of which corresponds to each pause sign are included; the dictionary in which elements each of which corresponds to each pause sign are included, a first recognition means which detects the position of the pause sign in the inputted speech; a decision means which decides an utterance section by making the input speech of the head position of the input speech which are positioned before the detected pause sign or the input speech which are positioned on and after the previous pause sign one sentence; and a second recognition means which performs speech recognition to the decided utterance section. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a continuous input speech recognition apparatus and a continuous input speech recognition method that do not require speech segmentation at the speech input stage for continuous input speech to be recognized.

  In general, in speech recognition, an acoustic model that retains phoneme features, a language model that retains speech content, and a dictionary that maintains the correspondence between phonemes and morphemes (words) and bridges the acoustic model and language model Use and to recognize the input speech.

  Conventionally, a language model holds a large number of example sentences in a recognition target field in parallel, and comparison with input speech has been performed between these parallel sentences. Therefore, the process of dividing the continuous input speech to be recognized into sentences, that is, speech segmentation, is performed manually or automatically at the speech input stage. In this voice segmentation, a sentence break is determined by detecting a silent section based on the power of an input voice signal or the like, and voice recognition is performed on the detected voice section.

On the other hand, for acoustic models, a technique for improving recognition performance by simultaneously considering the type of background noise and SNR (Signal to Noise Ratio) has been disclosed (see Non-Patent Document 1, for example).
Zhang, Sugimura, Furui: "Noise adaptation by tree-structure clustering considering both noise type and SNR", Acoustical Society Autumn Meeting, 2003

  As described above, conventionally, since a sentence break is determined by detecting a silence interval based on the power of the input speech signal, the interval detection becomes difficult in a noisy environment, and the recognition performance is greatly reduced. There was a problem. In particular, the conventional technique cannot cope with an environment in which noise fluctuates, that is, in a case where the noise type and SNR fluctuate.

  The present invention has been proposed in view of the above-described conventional problems, and the object of the present invention is to eliminate the need for speech segmentation at the speech input stage for continuous input speech to be recognized, and to achieve a recognition rate. It is an object to provide a continuous input speech recognition apparatus and a continuous input speech recognition method with high accuracy.

  In order to solve the above-described problems, in the present invention, as described in claim 1, an acoustic model that retains phoneme characteristics, a language model that retains speech content, and a correspondence between phonemes and morphemes are retained. An apparatus for recognizing input speech using a dictionary, the language model in which a plurality of sentences are connected by a pose symbol, the acoustic model including features of phonemes corresponding to the pose symbol, and the pose The dictionary including elements corresponding to symbols; first recognition means for detecting the position of the pose symbol in the input speech; and the leading or previous of the input speech before the detected pose symbol. A determination means for determining an utterance section using the input speech after the pause symbol as one sentence and a second recognition means for performing speech recognition on the determined utterance section are provided.

  According to a second aspect of the present invention, when the pause symbol is not detected, the second recognition means can perform voice recognition for the input voice having a predetermined length.

  In addition, as described in claim 3, a tree structure noise superimposed speech model that simultaneously considers the type of background noise and SNR is used as the acoustic model, and an optimal model is selected from the tree structure noise superimposed speech models. It is possible to select and perform voice recognition.

  In addition, as described in claim 4, using the recognition result by the first recognition means, linearly transform the parameters of the tree-structured noise superimposed speech model so as to maximize the likelihood, and after the linear transformation Speech recognition by the second recognition means can be performed using the tree structure noise superimposed speech model.

  According to a fifth aspect of the present invention, the processes of the first recognition unit, the determination unit, and the second recognition unit are repeatedly applied to the input speech that arrives continuously. Can do.

  In addition, as described in claim 6, input speech recognition is performed using an acoustic model that retains phoneme characteristics, a language model that retains speech content, and a dictionary that retains correspondence between phonemes and morphemes. A method of creating the language model as a combination of a plurality of sentences with pose symbols in a prior learning stage, and creating the acoustic model including features of phonemes corresponding to the pose symbols A step of creating the dictionary including an element corresponding to the pose symbol, and a first recognition step of detecting the position of the pose symbol in the input speech in the recognition step of the input speech. A determination step for determining an utterance section using the input speech at the beginning of the input speech or after the previous pause symbol as a sentence before the pause symbol, and the determined speech section And second recognition process to perform recognition can be configured as a continuous input speech recognition method comprising.

  In the continuous input speech recognition apparatus and continuous input speech recognition method of the present invention, the language model is a plurality of sentences connected with pause symbols, and the acoustic model and the dictionary are reflected with pause symbols, When recognizing the input speech, the pause symbol is recognized as a sentence break and recognized, so there is no need for speech segmentation at the speech input stage for continuous input speech to be recognized and the influence of noise. The recognition rate can be greatly increased.

  Preferred embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 is a flowchart showing processing of a continuous input speech recognition method according to an embodiment of the present invention.

  In FIG. 1, in the prior learning stage, first, a language model holding the utterance content is created (step S1). Specifically, as shown in FIG. 2, the initial symbol Start and the end symbol End are paused for the language model creation text T1 in which a plurality of example sentences are delimited by the initial symbol Start and the end symbol End and written in parallel. The language model creation text T2 in which a plurality of sentences are connected by pause symbols is created by rewriting the symbol LP, and a language model is created based on the language model creation text T2.

  Next, returning to FIG. 1, a dictionary holding correspondence between phonemes and morphemes (words) is created (step S2). Here, a dictionary including elements corresponding to the pose symbols is created.

Next, an acoustic model holding the phoneme features is created (step S3). Here, an acoustic model including the phoneme features corresponding to the pose symbol is created. In the present embodiment, a tree-structured noise-superimposed speech model that simultaneously considers the type of background noise and SNR is used as the acoustic model. As shown in FIG. 3, the tree-structure-noise-superimposed speech model is provided with a plurality of nodes N _{1 to} N _k branched into a tree structure under the node N ₀ of the root. It retains phoneme features corresponding to the types and various SNRs, and functions individually as an acoustic model. The upper node has a characteristic obtained by averaging the characteristics of a plurality of lower nodes. Therefore, by selecting an appropriate node, high recognition adapted to the type of noise and various SNRs can be performed.

  Next, returning to FIG. 1, in the recognition stage, first, the maximum likelihood is obtained from the tree-structured noise superimposed speech model for an input speech of a certain length (an appropriate value is set according to the environment). Thus, the process of selecting the optimum model (node) and the first recognition process of detecting the position of the pause symbol in the input speech are performed (step S4).

  Next, a process of determining an utterance section is performed with the input speech before the detected pause symbol and at the beginning of the input speech or after the previous pause symbol as one sentence (step S5). When no pause symbol is detected, the entire input speech of a certain length that is the processing target is used as it is as an utterance section.

  Next, using the recognition result obtained by the first recognition process, linear conversion is performed so that the parameters of the tree-structured noise-superimposed speech model have the maximum likelihood (step S6). Although the recognition rate slightly decreases, this linear conversion can be omitted.

  Next, a second recognition process is performed on the determined speech section using the linearly converted tree-structured noise superimposed speech model (step S7).

  These processes are repeatedly applied to the input voice.

  In this way, the language model is composed of a plurality of sentences connected by pose symbols, and the acoustic model and dictionary reflect the pose symbols, and when the input speech is recognized, the pose symbols are determined as sentence breaks. Since recognition is performed, voice segmentation at the voice input stage is not required for continuous input voice to be recognized. In addition, since the sentence can be appropriately extracted as an utterance section without being affected by noise, the recognition rate can be significantly increased.

  In addition, a tree-structured noise-superimposed speech model is used as the acoustic model, and an optimal model is selected and linear conversion is performed using the first recognition result. Therefore, even for input speech that includes various types of noise, The recognition rate can be further improved.

  Next, FIG. 4 is a diagram showing a speech recognition algorithm, (a) showing continuous input speech recognition of the present invention, and (b) showing conventional speech recognition for comparison.

  In FIG. 4A, an optimal model (node) is selected from the acoustic model (sound model with superimposed noise of tree structure) M1 for the input speech IN having a certain length (step S11), and the selected acoustics are selected. The first speech recognition is performed using the model M1, the language model M2, and the dictionary D (step S12).

  Next, the speech section is determined from the position of the pause symbol in the input speech IN detected by the first speech recognition (step S13), and linear transformation adaptation of the acoustic model M1 is performed (step S14). A second speech recognition is performed on the speech section using the linearly converted content of the acoustic model M1, the language model M2, and the dictionary D (step S15), and a result OUT is obtained.

  In FIG. 4B, conventionally, section detection (speech segmentation) is performed on the input voice IN having a certain length by signal processing based on the power of the input voice signal and the like (step S21). The result OUT is obtained by performing speech recognition (step S22). If the accuracy of the section detection is deteriorated due to background noise included in the input speech IN, the recognition rate is directly reduced. In the present invention, however, speech segmentation is performed. Without performing this, it is possible to prevent the recognition rate from being lowered due to noise.

  Next, FIG. 5 is a diagram showing a configuration example of the continuous input speech recognition apparatus of the present invention.

  In FIG. 5, the continuous input speech recognition system includes a model / dictionary storage unit 1 that holds a language model, a dictionary, and an acoustic model (a tree-structured noise superimposed speech model), and features that analyze the input speech IN and convert it into a feature vector. An extraction unit 2 and a model selection unit 3 that selects an optimal model (node) from the acoustic model of the model / dictionary storage unit 1 according to the feature vector output from the feature extraction unit 2 so as to maximize the likelihood. The speech recognition unit 4 that performs the first speech recognition on the input speech IN converted to the time-series data of the feature vector, and the speech segment determination that determines the speech segment based on the detection result of the pause symbol by the speech recognition unit 4 Unit 5, model linear transformation adaptation unit 6 that performs linear transformation so that the likelihood of the acoustic model is maximized based on the recognition result by speech recognition unit 4, and the speech of which input speech IN is determined To between, and a speech recognition unit 7 for speech recognition for the second time using a linear transformed acoustic models, and a recognition result storage unit 8 for storing the recognition result. The voice recognition unit 4 and the voice recognition unit 7 are described separately because of different processing phases, but are the same as the program (engine).

  Next, FIG. 6 is a diagram showing an example of a recognition rate experimental result examined using noise speech to be recognized in order to confirm the effect of the present invention. The acoustic model used in the experiment is an unspecified speaker context-dependent phoneme HMM (Hidden Markov Model) in which state sharing is performed by tree-based clustering. As the feature quantity, a total of 25 dimensions of MFCC (Mel Frequency Cepstrum Coefficient) 12 dimensions, its differential 12 dimensions, and first derivative of logarithmic power was used. FIG. 6 shows the correct word accuracy (target value) when a correct utterance section is designated, and the correct word accuracy (present invention) according to the embodiment of the present invention. From this result, in the embodiment of the present invention, it can be seen that the result is the same as when the correct utterance section is designated.

  The present invention has been described above by the preferred embodiments of the present invention. While the invention has been described with reference to specific embodiments, various modifications and changes may be made to the embodiments without departing from the broad spirit and scope of the invention as defined in the claims. Obviously you can. In other words, the present invention should not be construed as being limited by the details of the specific examples and the accompanying drawings.

It is a flowchart which shows the process of the continuous input speech recognition method concerning one Embodiment of this invention. It is a figure which shows the example of preparation of a language model. It is a figure which shows the structural example of a tree structure noise superimposition speech model. It is a figure which shows the algorithm of speech recognition. It is a figure which shows the structural example of the continuous input speech recognition apparatus of this invention. It is a figure which shows the example of the experiment result of a recognition rate.

Explanation of symbols

T1 to T2 Language model creation text IN input speech OUT result M1 acoustic model M2 language model D dictionary 1 model / dictionary storage unit 2 feature extraction unit 3 model selection unit 4 speech recognition unit 5 utterance interval determination unit 6 model linear transformation adaptation unit 7 Speech recognition unit 8 Recognition result storage unit

Claims (6)

An apparatus for recognizing input speech using an acoustic model that retains phoneme features, a language model that retains speech content, and a dictionary that retains correspondence between phonemes and morphemes,
The above language model in which multiple sentences are connected with pause symbols,
The acoustic model including the phoneme features corresponding to the pose symbol;
The dictionary including elements corresponding to the pose symbols;
First recognition means for detecting the position of the pose symbol in the input speech;
A determination means for determining an utterance section using the input speech before the detected pause symbol and at the beginning of the input speech or after the previous pause symbol as one sentence;
A continuous input speech recognition apparatus comprising: a second recognition unit that performs speech recognition on the determined utterance section.   The continuous input speech recognition apparatus according to claim 2, wherein when the pause symbol is not detected, speech recognition is performed by the second recognition means for the input speech having a predetermined length. As the acoustic model, a tree-structured noise superimposed speech model that simultaneously considers the type of background noise and SNR is used.
The continuous input speech recognition apparatus according to any one of claims 1 and 2, wherein speech recognition is performed by selecting an optimal model from the tree structure noise superimposed speech model.   Using the recognition result obtained by the first recognition means, the tree-structured noise-superimposed speech model parameters are linearly transformed to maximize the likelihood, and the tree-structured noise-superposed speech model after the linear transformation is used to perform the first conversion. The continuous input speech recognition apparatus according to claim 3, wherein speech recognition is performed by the recognition means of 2.   5. The process according to claim 1, wherein the processes of the first recognition unit, the determination unit, and the second recognition unit are repeatedly applied to the input speech that arrives continuously. The continuous input speech recognition apparatus as described. A method for recognizing input speech using an acoustic model that retains phoneme features, a language model that retains speech content, and a dictionary that retains correspondence between phonemes and morphemes,
In the pre-learning phase,
Creating the language model as a plurality of sentences connected by pause symbols;
Creating the acoustic model including phoneme features corresponding to the pose symbol;
Creating the dictionary including elements corresponding to the pose symbols;
In the input speech recognition stage,
A first recognition step of detecting the position of the pose symbol in the input speech;
A determination step of determining an utterance section using the input speech before the detected pause symbol and at the beginning of the input speech or after the previous pause symbol as one sentence;
A continuous input speech recognition method comprising: a second recognition step of recognizing the determined utterance section.

Images