JPH05128286A - Keyword spotting system by neural network - Google Patents

Abstract

PURPOSE:To establish the effective constitution method and learning method of a neural network, regarding the spotting system of keyword sound having a wide versatility. CONSTITUTION:When sound is inputted in a sound input part 5, the feature analysis of sound such as a FFT(fast Fourier transform) is performed in a feature extraction part 1. As the analysis is a keyword spotting, the segment processing of sound section especially by sound wave power is not performed. In a learning mode 6, a discrimination learning is performed for keywords and sound other than keywords (including noise) by using a neural network in a neural net learning part 2. In a recognition mode 7, a spotting is performed for keywords by using the learned neural network in a neural network recognition part 3. The spotting result is outputted in a keyword detection part 4.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a voice keyword spotting method using a neural network. For example, it is generally applied to a voice recognition device using a neural network.

[0002]

2. Description of the Related Art As a known document describing the prior art of the present invention, for example, "Study of syllable spotting by time-delayed neural network" (Sawai, Weibel, Shikano, Acoustical Society of Japan 2-P-11, PP223 -224, October 1988). This document describes a time-delayed neural network (Time-D
elay Neural Network (TDNN) is used for spotting single syllables in Japanese using a neural network. Taking "BA" as a syllable example,
This and "DA", "GA", which are considered to be particularly erroneous
The discriminative learning with five syllables of "PA", "TA", and "KA" was performed using the error back-propagation method (backpropagation). As a result, "BA" in the Japanese word speech was obtained with good performance. I was able to spot and suppress all syllables except "BA." However, in the prior art,
Neural networks were only applied to the very limited target of Japanese monosyllabic spotting.

[0003]

[Purpose] The present invention has been made in view of the above-mentioned circumstances, and establishes an effective neural network configuration method and learning method for a keyword speech spotting method with great versatility. The objective is to construct an effective learning method of a neural network even when the included voice includes noise. That is, in the keyword spotting method using a neural network, since there are innumerable words other than the keywords for the keywords, it is necessary to efficiently select the words (in contrast, the minimal pair of phonemes for the keywords is used). (Select words that differ only in one phoneme.) Also, for the purpose of reducing the scale of the neural network, easiness of learning, and easiness of selecting a keyword and a word pair other than the keyword, 1 for each keyword. If one group of keywords can be integrated and integrated into one neural network as a coping method when the number of keywords increases, a more compact keyword spotting method can be realized, and the keyword speech Other than, There is noise outside the voice, but by using the learning ability of the neural network to distinguish the voice other than the keyword including the noise from the keyword, and as a learning method of the neural network, noise was superimposed at once. Sometimes it is not easy to distinguish between speech and keywords.Therefore, at first, learning is started from a keyword pair with no noise and other word pairs, and it becomes stable by gradually adding noise. The purpose is to control the learning of the neural network.

[0004]

In order to achieve the above object, the present invention provides (1)
A feature extraction unit that inputs a voice waveform including a predetermined keyword and performs feature analysis, a neural network learning unit that learns a neural network using the feature amount extracted by the feature extraction unit, and the neural network learning A neural network that has a recognition unit by a neural network that performs keyword recognition using a neural network that has already been learned by the unit, and a keyword detection unit that outputs the keyword recognized by the recognition unit, and that distinguishes between the keyword and non-keywords. In the keyword spotting method of performing keyword spotting using, a keyword is detected by using a neural network that learns pairs with a plurality of words having a minimum phoneme pair for the keyword, and (2) one keyword Against Then, one neural network is made to correspond, the value "1" of the output unit of the neural network is made to correspond to a keyword, and "0" is made to correspond to a word other than the keyword, and the error back propagation method (back propagation method) is applied. Learning by using (3)
Use a single neural net for multiple keywords and have an output unit for each keyword,
Furthermore, (4) noise is superposed on the word other than the keyword, and the neural net is learned so that the keyword and the voice other than the keyword on which the noise is superimposed can be distinguished, and (5) the above ( In 4), the learning of the neural network is sequentially advanced while gradually adding the noise superposition. Hereinafter, description will be given based on examples of the present invention.

FIG. 1 is a block diagram for explaining an embodiment of a keyword spotting system by a neural network according to the present invention. In the figure, 1 is a feature extraction unit, 2 is a neural network learning unit, and 3 is a neural network recognition. Part 4, keyword detection unit, 5 voice input unit, 6 learning mode, 7 recognition mode. First, when a voice is input from the voice input unit 5, the feature extraction unit 1 performs a voice feature analysis such as FFT (Fast Fourier Transform). Since the keyword spotting is performed, the voice segment is not particularly cut out by the voice waveform power. In the learning mode 6, the neural network learning unit 2 performs discrimination learning of a keyword and a voice (including noise) other than the keyword using a neural network. In the recognition mode 7, the neural network recognition unit 3 performs keyword spotting using a learned neural network. The keyword detection unit 4 outputs the spotting result.

FIG. 2 is a diagram for explaining the arrangement of keywords and voices other than the keywords in the feature vector space. In general, the feature space is a multidimensional space. 21
Is a local space in which keywords are distributed, and 21a to 21g are spaces in which voices (including noise) other than the keywords are distributed. 22a to 22d are hyperplanes for distinguishing between keywords and non-keywords. This is the neural network learning unit 2 of FIG. 1 and uses the error back propagation method (back propagation).
Is formed by. Here, the keyword is, for example, "θNs
"ei (speech)" means that only one phoneme is different from the words other than the keywords, so-called minimal pair words (there is no need to stick to meaningful words,
It is desirable that it is a meaningful word because it is easy to speak). For example, "θNseN (hot spring)" or "KoNsei (mixed)"
And so on. As a method of creating these words, it is also possible to artificially create words in which one phoneme is randomly or randomly replaced and select a meaningful word from these words. Words that differ from these keywords only in one phoneme occupy spatial positions 21a to 21g adjacent to the keywords in the multidimensional space formed by the feature amounts of the entire word pattern. Therefore, the 22
If the discrimination boundary surfaces such as a and 22b are formed by learning, all words other than the minimal pair can be automatically discriminated from the keywords. As a result, the keyword can be spotted from the voice by distinguishing it from words other than the keyword.

FIG. 3 is a diagram showing an embodiment of a neural network (NN) for keyword spotting. In FIG.
1 is an input layer, 32 is an intermediate layer, 33 is an output layer (unit), 31a is a coupling portion between the input layer and the intermediate layer, and 32a is a coupling portion between the intermediate layer and the output layer. In the figure, three-layer NN
However, the number of layers may be 4 or more. The figure shows the case where the output layer has a single output unit. In this case, the keyword is "1" and the keyword other than the keyword is "0". In this example, 1 for each keyword
This is a case where one NN (and one output unit) is assigned. In the case of a plurality of keywords (m is set), the number of output units in the output layer 33 is m. At this time, for the i-th keyword,

[0008]

[Table 1]

The output unit is learned so as to fire. Teach all values except 0 to keywords.

FIG. 4 is an explanatory diagram of the keyword spotting method. 10 is a voice waveform including a keyword, 13 is an FFT obtained by analyzing the voice waveform 10.
The output, A, is the keyword spotting neural network shown in FIG. The neural network 3 is sequentially scanned from the beginning of the voice, and the presence of the keyword can be confirmed when the output unit 33 fires.
In the figure, it can be seen that the output unit 33 is firing when the keyword presence section is reached.

[0011]

As is apparent from the above description, the present invention has the following effects. (1) Effect corresponding to claim 1: Since words of a minimal pair are used as word sounds other than keywords, the learning amount of the neural network can be minimized by constructing the necessary minimum word pairs. In addition, words not used for learning can be automatically rejected by distinguishing them from keywords. (2) Effect corresponding to claim 2: Since one neural network is associated with one keyword,
When the number of keywords is relatively small, the network scale can be reduced, learning can be easily performed, and words other than keywords can be easily selected. (3) Effect corresponding to claim 3: When the number of keywords is relatively large, a keyword spotting device with a more compact neural network as a whole can be realized by integrating a plurality of keywords into one neural network. .. (4) Effect corresponding to claim 4: Since noises are learned as voices other than keywords, noises can be automatically removed at the time of keyword detection. (5) Effect corresponding to claim 5: When noise is added to a voice other than a keyword, learning of the neural network is first started from a state without noise, and noise is gradually superimposed. , The learning of the neural network can proceed stably.

[Brief description of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a keyword spotting method using a neural network according to the present invention.

FIG. 2 is a diagram for explaining an arrangement of keywords and phonemes other than the keywords in a feature vector space.

FIG. 3 is a diagram showing an embodiment of a neural net (NN) for keyword spotting.

FIG. 4 is an explanatory diagram of a keyword spotting method.

[Explanation of symbols]

1 ... Feature extraction unit, 2 ... Neural network learning unit, 3 ... Neural network recognition unit, 4 ... Keyword detection unit,
5 ... voice input unit, 6 ... learning mode, 7 ... recognition mode.

Claims (5)

[Claims] 1. A feature extraction unit for inputting a speech waveform containing a predetermined keyword and performing a feature analysis, and a neural network learning unit for learning a neural network using the feature amount extracted by the feature extraction unit. A keyword recognition unit that recognizes a keyword using a neural network that has been learned by the neural network learning unit, and a keyword detection unit that outputs the keyword recognized by the recognition unit. In a keyword spotting method that performs keyword spotting using a neural network that distinguishes between words, a neural network characterized by detecting a keyword using a neural network that learns pairs with a plurality of words that have a minimum phoneme pair for a keyword. Network keyword spotting method. 2. A neural network is associated with one of the keywords, a value "1" of an output unit of the neural network is associated with the keyword, and "0" is associated with a word other than the keyword, The keyword spotting method using a neural network according to claim 1, wherein learning is performed using an error backpropagation method. 3. The keyword spotting method according to claim 1, wherein a single neural net is used for a plurality of keywords, and an output unit for each keyword is provided. 4. The neural network according to claim 1, wherein noise is superposed on a word other than the keyword, and the neural network is learned so that the keyword and the voice other than the keyword on which the noise is superimposed can be distinguished. Network keyword spotting method. 5. The keyword spotting method by a neural network according to claim 4, wherein learning of the neural network is sequentially advanced while gradually adding the noise superposition.