JPS6329278B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an audio device, and more particularly to an audio device that registers audio in a registration mode and then recognizes the registered audio in a recognition mode.

2. Description of the Related Art Speech devices that extract speech features and perform recognition by comparing the extracted speech features with the features of subsequently input speech have appeared and are being put into practical use.
There are two types of voice devices: one that recognizes the voice of a specific person and one that recognizes the voice of an unspecified speaker, but the former is currently the best in terms of recognition time, recognition rate, and device configuration. Most of them are for specific speakers.

The present invention aims to improve the usability of such an audio device, and will be described in detail below with reference to the drawings.

1 is a microphone that converts audio into an electrical signal, and 2 is a feature extraction circuit that extracts the features of the audio electrical signal obtained from this microphone, which consists of a zero-cross detection circuit, an audio spectrum extraction circuit, an audio region detection circuit, etc. It's on. Reference numeral 3 denotes a mode changeover switch for switching between a voice registration mode and a voice recognition mode, and 4 a first feature pattern memory connected to the registration mode side of the mode changeover switch 3 via a first input gate group 5. in,
The voice feature patterns extracted by the feature extraction circuit 2 are stored. 6 is also a second feature pattern memory connected to the registration mode side of the mode changeover switch 3 via the second input gate group 7, and is connected to the feature extraction circuit 2.
The extracted voice feature patterns are stored.
Note that these feature pattern memories 4 and 6 each have a capacity to store speech feature patterns of, for example, about 8 words, and the writing and reading timing of the feature patterns to each memory 4 and 6 is determined by the following timing. For example, T ₁ to _{T 8} and T ₉ to T ₁₆ ,
They are not performed in the same way and are shifted. Reference numeral 8 denotes a buffer memory connected to the recognition mode side of the mode changeover switch 3, in which the characteristic pattern of the unknown voice from the characteristic extraction circuit 2 is temporarily stored. 9,10
are first and second output gate groups provided on the output side of the first and second feature pattern memories 4 and 6,
11 is a recognition circuit that compares and recognizes the registered pattern obtained through these output gate groups 9 and 10 and the unknown voice pattern from the buffer memory 8; 12;
is a speaker selection switch that supplies switching signals to each input/output gate group 5, 7, 9, and 10, and selects three types of speakers: first speaker A, second speaker B, and specific speaker Z. Speaker selection is performed. Next, the internal configuration of each input/output gate group 5, 7, 9, and 10 will be explained. The first input/output gate group 5, 9 includes a first
OR gates 20 and 22 that take the logical sum of speaker A and specific speaker Z; AND gates 21 and 23 that take the logical product of the logical sum output and the voice feature pattern;
It consists of Second input/output gate group 7, 10
is an OR that is logically ORed with the second speaker B and the specific speaker Z.
gates 24 and 26, and AND gates 25 and 27 that take the logical product of the logical sum output and the voice pattern;
It consists of

First, we will explain the operation of voice registration. When the mode changeover switch 3 is set to the registration mode side and the registered speaker is selected as the first speaker A,
The characteristic pattern of the first speaker's voice obtained through the microphone 1 and the characteristic extraction circuit 2 passes through the first input gate group 5 and is stored in the first characteristic pattern memory 4 for up to eight words. Further, when the second speaker B is selected, the second input gate group 7 is opened and the voice feature pattern of the second speaker is stored in the second feature pattern memory 6 for up to eight words. Further, when a specific speaker Z is selected by the speaker selection switch 12, both input gates 5 and 7 are opened, and the feature pattern of the specific speaker is introduced into the first and second feature pattern memories 4 and 6. As mentioned above, since the writing timings of the feature patterns to each memory 4 and 6 are shifted, a feature pattern of 8+8, ie, a maximum of 16 words, is stored in both memories 4 and 6.

The recognition operation using the voice registered in this way will be explained next. Mode change switch 3
If the first speaker is to be recognized by switching to the recognition mode side, the first speaker A is designated by the speaker selection switch 12. In this state, when the first speaker utters an unknown voice into the microphone 1, its features are extracted by the feature extraction circuit 2, and the buffer memory 8
temporarily stored. Then, the temporarily stored characteristic pattern of the unknown voice and the characteristic pattern of the registered voice registered in the first characteristic pattern memory 4 are compared in the comparison recognition circuit 11, and the unknown characteristic pattern is selected from among the registered patterns. It is recognized specifically by either. When the second speaker B is selected, the characteristic pattern of the unknown voice of the second speaker B is compared with the pattern registered in the second characteristic pattern memory 6 and recognized by the comparison recognition circuit 11.

Next, the selection switch 12 which is the most characteristic of the present invention
Let us consider the case where specific speaker Z is selected. In the voice registration mode, when specific speaker Z is selected with the voices of the first and second speakers registered in the first and second characteristic pattern memories 4 and 6, respectively, the voices of the first and second speakers are registered in the buffer memory 8. The temporarily stored unknown voice and the registered patterns of the voices of both the first and second speakers are compared in a comparison recognition circuit 11, and the most similar voice pattern among the registered patterns is selected to recognize the unknown voice. That is, in this state, the voices of the two speakers are recognized.

On the other hand, in the registration mode, a specific speaker's (8+
8) When a specific speaker is selected by the selection switch 12 with the word registered in both feature pattern memories 4 and 6, the content of the buffer memory 8 and the specific speaker's 16 in both feature pattern memories 4 and 6 are selected. The contents of the words are compared in a comparison recognition circuit 11, and unknown speech is selected from among the 16 words and recognized.

As is clear from the above description, the present invention has the following features:
The second speaker's voice is recognized individually, and when the voices of both speakers are registered and a specific speaker is selected and recognized, the unknown voice is recognized as the voice of both speakers. If you perform a recognition operation for a specific speaker, and select that specific speaker and perform recognition with the voice of a specific speaker registered, it is possible to recognize twice the number of voices normally. be. Therefore, it is possible to deal with a large number of speech recognition formats, increase the versatility of the speech recognition device, and at the same time, it is possible to obtain an extremely practical speech device with a small burden on hardware.

[Brief explanation of the drawing]

The figure is a block diagram showing the configuration of a voice device that implements the voice recognition processing method of the present invention, in which 2 is a feature extraction circuit, 4 and 6 are feature pattern memories, 8 is a buffer memory, 11 is a comparison recognition circuit, and 12 is a A speaker selection switch is shown.

Claims (1)

[Claims] 1. A voice recognition processing method that registers voice in a registration mode and then recognizes the registered voice in a subsequent recognition mode, which includes: a microphone that converts voice into an electrical signal; a feature extraction circuit for extracting the features of the audio electrical signal from the microphone; and a first and second feature extraction circuit capable of storing feature patterns of n words (n is an integer of 1 or more) extracted by the feature extraction circuit. 2, a buffer memory for temporarily storing the feature pattern of the input voice obtained from the feature extraction circuit, and recognition for comparing and comparing the contents of the buffer memory with the contents of the first and second feature pattern memories. a circuit, a mode switching means for switching between a voice registration mode and a voice recognition mode, and a speaker selection means for a first speaker, a second speaker, and a specific speaker; The feature pattern of n words of the speaker's voice is stored in the first feature pattern memory, the feature pattern of n words of the second speaker's voice is stored in the second feature pattern memory, and the feature pattern of 2n words of the specific speaker is stored in the first feature pattern memory. 1 and 2 are respectively stored in the feature pattern memory, and when the first speaker is selected in the recognition mode, the content introduced in the buffer memory and the content stored in the first feature pattern memory are When the second speaker is selected, the recognition circuit compares and recognizes the contents introduced in the buffer memory and the contents stored in the second feature pattern memory, and , the second speaker's voice feature pattern as the first,
When a specific speaker is selected in the recognition mode with the data stored in the second feature pattern memory, the contents of the buffer memory and the contents of both feature pattern memories are compared and recognized to recognize the voices of the two speakers. When a specific speaker is selected in the recognition mode with the feature pattern of 2n words stored in both feature memories, the content of the buffer memory and the content of both feature memories are compared and the 2n word feature pattern is stored in both feature memories. A speech recognition processing method characterized by performing speech recognition of words.