JPH04128898A - Voice recognition device - Google Patents

Abstract

PURPOSE:To easily increase a vocabulary and speakers to be recognized by generating the standard pattern of an input voice. CONSTITUTION:When an input voice is used for registration, a switch 12 is switched to the side of a standard pattern generation part 14 and the standard pattern of the vocabulary is generated for a voice pattern converted by a signal conversion part 11 by using data stored in a memory part 15. When the input voice is used for discrimination, the switch 12 is switching to the side of a similarity calculation part 13, the similarity between the registered standard pattern and input voice pattern is calculated and a decision part 17 outputs the standard pattern with the highest similarity. Standard patterns corresponding to three kinds of vocabulary of one speaker are registered in a memory part 16 for the standard patterns, but when a sufficient recognition result is obtained by the voicing of a different speaker, new voice registration is not performed.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Field of Industrial Application) This invention registers a speaker's voice as a standard pattern in advance, and calculates the similarity between an input voice pattern and this standard pattern. The present invention relates to a speech recognition device that recognizes input speech.

(Prior Art) A speech recognition device calculates the degree of similarity between a standard pattern for speech recognition and an input speech pattern, and recognizes the input speech based on the calculation result.

A speech recognition device that recognizes input speech in this way is
Speech recognition devices are generally classified into speech recognition devices for non-specific speakers and speech recognition devices for specific speakers. In the former speech recognition device for unspecified speakers, a word to be recognized is determined in advance, and a standard pattern for speech recognition is created using the speech of the word by a number of speakers.

In the latter speech recognition device for a specific speaker, g49 that is desired to be recognized by the specific speaker used is registered in advance.

(Problem to be Solved by the Invention) Incidentally, the former speech recognition device for unspecified speakers does not require vocalization during speech recognition, but has a problem in that it can only recognize speech from vocabulary that has been prepared in advance. .

In addition, unlike the former, the latter type of speech recognition device for specific speakers does not have any restrictions on the speakers to be recognized, but there is a problem that recognition performance deteriorates when a person other than the registered speaker uses the speech recognition device. There is. Therefore, in order for a person other than the speaker who has registered the standard pattern to use the recognition device, it is necessary to newly register the standard pattern. However, as the number of speakers using the speech recognition device increases, a problem arises in that a proportionally large amount of memory is required to store standard patterns.

This invention was made in consideration of the above circumstances, and its purpose is to easily increase the number of word searches, easily increase the number of speakers to be recognized, and increase the number of speakers to be recognized. To provide a speech recognition device that does not involve an increase in standard pattern capacity in proportion to the number of speakers even when the number of speakers increases.

[Structure of the Invention] (Means for Solving the Problems and Their Effects) The speech recognition device of the present invention includes a standard pattern creation section that creates a standard pattern for registration from input speech, and a standard pattern creation section that creates a standard pattern for registration from input speech. a first memory section that stores data used when creating standard patterns for the above, and a second memory section that stores standard patterns corresponding to various voices in advance and also stores standard patterns created by the pattern creation section. The present invention is characterized by comprising a memory section, and a similarity calculation section that calculates the similarity between the standard pattern stored in the second memory section and the speech input pattern for recognition.

In the above invention, when a voice with the same vocabulary as the standard pattern previously stored in the second memory section is uttered by a different speaker at the time of registration, the pattern creation section matches the input voice with the first one.
A new standard pattern for registration is created from the data stored in the second memory section, and is re-stored in the area where the standard pattern of the corresponding vocabulary is previously stored in the second memory section.

Additionally, if the voice of a new word search for which the standard pattern is not stored in the second memory section at the time of registration is uttered for the first time by a different speaker, the pattern creation section will combine the input voice with the standard pattern stored in the first memory section. A standard pattern for registration is created from the existing data and newly stored in the second memory section.

Further, the speech recognition device of the present invention includes a memory section that stores in advance standard patterns corresponding to various voices, and creates a new standard pattern using the input voice and the standard patterns stored in the memory section. The present invention is characterized by comprising a pattern creation section that causes the pattern to be re-stored in the memory section, and a similarity calculation section that calculates the similarity between the standard pattern stored in the memory section and the voice input pattern for recognition.

In the above invention, when a voice with the same number of words as the standard pattern stored in the memory section is uttered by a different speaker at the time of registration, the pattern creation section compares the input voice with the standard pattern stored in the memory section. A new standard pattern is created from this and stored again in the memory section.

(Examples) Hereinafter, the present invention will be explained by examples with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a first embodiment of a speech recognition device according to the present invention. The input voice is converted into a voice pattern in the signal converter 11.

The voice pattern converted here is selectively supplied to the similarity calculation section 13 and the standard pattern creation section 14 via the switch 12. In the standard pattern creation section 14, an input voice pattern and a memory section 15 for standard pattern creation data are stored.
A standard pattern for registration is created from the input voice pattern using the standard pattern creation data stored in . Memory section 15 for the above standard pattern creation data
stores core data for standard pattern creation, which is created in advance using sounds of specific words uttered by a plurality of speakers. The standard pattern created by the standard pattern creation section 14 is stored in a standard pattern memory section 16.
and stored in a predetermined area.

On the other hand, the similarity calculation unit 13 calculates the similarity between the input voice pattern and the standard pattern stored in the standard pattern memory unit IB, and the calculation result is sent to the determination unit 1.
Sent to 7. Based on the calculation result of the degree of similarity calculated between each standard pattern and the input speech pattern, the determination unit 17 outputs a recognition result indicating that the input voice has the vocabulary corresponding to the standard pattern with the highest degree of similarity. be done.

Next, the operation of the apparatus having the above configuration will be explained. First, when a voice is input, the switch 12 is changed depending on whether the input voice is used for recognition or registration. Note that this switching operation of the switch 12 is performed under the control of a control circuit (not shown). When input audio is used for registration, the switch 12 is switched to the standard pattern creation section 14 side.

Therefore, the audio pattern converted by the signal converter 11 is supplied to the standard pattern generator 14. In the standard pattern creation unit 14, a standard pattern of the corresponding word search is created using the data stored in advance in the standard pattern creation data memory unit 15 and the converted voice pattern, and then the standard pattern is stored in the standard pattern creation data memory 15. 16.

On the other hand, when the input voice is used for identification, the switch 12
is switched to the similarity calculation unit 13 side. In this case, the audio pattern converted by the signal conversion section 11 is supplied to the similarity calculation section 13. Then, the similarity calculation unit 13 calculates the similarity between the input speech pattern and the standard pattern corresponding to some words registered (stored) in advance, and then the determination unit 17 calculates the similarity between the input speech pattern and the standard pattern corresponding to some words registered (stored) in advance. A recognition result is output indicating that the input voice corresponds to a high standard pattern.

Next, the operation when a plurality of speakers use the device of the above embodiment will be explained. In addition, in the standard pattern memory section 1B,
As shown in Figure 2, a speaker uses three types of words *w1.
．． w2. Standard patterns Wl, W2. corresponding to w3. W3
is registered (stored) in advance. In this state, a different speaker than when registering is using the above three types of words.
, w2w3 is uttered for recognition, and if sufficient recognition results are obtained, no new voice is registered. However, if it is necessary to register a new rhyme or to register the user's voice for a phrase whose recognition performance is not satisfactory, new registration is performed. For example, as shown in FIG. 2, speaker A pronounces w2. When the voice w4 is uttered, the standard pattern creation unit 14 creates standard patterns W2' and W4 for each common word. The created standard pattern is sent to the standard pattern memory unit 16, but the standard pattern W2' corresponding to the word W2 of speaker A is stored again in the area where W2 was previously stored, and is stored in the new dielectric W4. The corresponding standard pattern W4 is stored in a new area of the standard 4 pattern memory section 16. In this state, a speaker B different from the above speaks the four types of words w 1 , w
2. w 3. When a voice corresponding to w4 is uttered to perform recognition, if a sufficient recognition result is obtained, no new voice is registered.

However, if it is necessary to additionally register a new dielectric voice or to register a voice for a word with low recognition performance for speaker B, a new registration is performed in the same manner as above. For example, the second
As shown in the figure, when speaker B utters the words $1wl and w5, the standard pattern creation section 14 creates standard patterns Wl' and W5 for the respective dielectrics. The created standard pattern is sent to the standard pattern memory unit 16, and a standard pattern Wl' corresponding to the word W1 of speaker B is sent to the standard pattern memory unit 16.
is re-stored in the area where Wl was previously stored,
The standard pattern W5 corresponding to the new word W5 is stored in a new area of the standard pattern memory section 16. Therefore, in the end, the standard pattern W4.corresponding to the dielectrics W4 and w5 is different from the original one. W5 is added or dielectric w l
, w2 is an updated standard pattern Wl'W2' that includes the characteristics of multiple speakers.

As described above, in the speech recognition device of the above embodiment, as the number of registered speakers for the same vocabulary increases, the characteristics of the speakers included in the standard pattern stored in the standard pattern memory section 16 change to those of a large number of speakers. Therefore, it is similar to a speech recognition device for unspecified speakers, and the recognition rate for unregistered speakers is improved.

In addition, as the standard pattern stored in the standard pattern memory section 16 changes as a large number of speakers register, it is possible to use a voice recognition system for multiple speakers at once, such as in a speech recognition device for unspecified speakers. There is no need to collect data or spend time creating standard patterns. Furthermore, when a plurality of speakers register standard patterns for the same vocabulary, a memory capacity proportional to the number of speakers is not required, and memory usage efficiency can be improved.

In the above embodiment, the standard pattern creation unit 14 uses the well-known dominant speaker equality method, dominant speaker identification method, smoothing/
It is possible to use a device that creates a standard pattern for registration from input audio using a differential pattern orthogonalization method, etc., and a mask ROM (
Read 0nly MeIlory), PROM
(Programmable ROM) EP
ROM (Erasable PROM), E2FROM
(Electrical Iy EPROM), C
D (CoIlpact Disk) device, FD (Flo
pl)y Disk) device and HD (Hard Disk)
) device, etc. can be used, and as the standard pattern memory section 16, a RAM (Random Access
Any recording medium on which data can be written can be used, such as a memory, an E2PROM, an FD device, an HD device, or an optical disk device.

The data stored in the standard pattern creation data memory section 15 is, for example, -L (Karhunen-
Loeve) expansion method, differential filter method, differential orthogonalization filter method, vector quantization method, pack propagation method, or the like. On the other hand, as a calculation method in the similarity calculation section 13, DP (
Dynamic Programming) matching (Mate old ng), composite similarity method, HMM ()Ild
The similarity between the input pattern and the standard pattern can be calculated using a similarity method using the denMarcon Model 1, a similarity method using fuzzy inference, or the like.

FIG. 3 is a block diagram showing the configuration of a second embodiment of the speech recognition device according to the present invention. This embodiment device differs from the device shown in FIG. 1 in that a memory section 15 for standard pattern creation data is not provided, and at the time of registration, the standard pattern creation section 14 uses the input audio pattern supplied via the switch 12. The two points are that a new standard pattern for registration is created using the standard pattern previously stored in the standard pattern memory section 16. Therefore, the standard pattern stored in the standard pattern memory section 16 changes sequentially each time a different speaker registers it, and it becomes a pattern that combines the characteristics of multiple speakers. The pattern will be lost.

However, compared to the embodiment apparatus shown in FIG. 1, the standard pattern creation data memory section 15 is not required, and the memory capacity can be reduced compared to the case shown in FIG.

FIG. 4 is a block diagram showing the configuration of a third embodiment of the speech recognition device according to the present invention. The difference between this embodiment device and the device shown in FIG. 1 is that the standard pattern memory section 1
6 and the standard pattern creation section 14 and the standard pattern creation data memory section 15, a switch 18 is newly provided.

In this embodiment, when the standard pattern creation section 14 creates a standard pattern for registration, the standard pattern creation data memory section 15
It is possible to select and use either the data stored in the standard pattern memory section 16 or the standard patterns stored in the standard pattern memory section 16. Therefore, when registering for the first time, the switch 18 is set to the memo 9 section 15 for standard pattern creation data.
side, and create a standard pattern for registration using the data stored in this standard pattern creation data memory section 15. After that, switch 18 to the standard pattern memo 9 section 16 side, and create the standard pattern. It is possible to register the voices of different speakers by sequentially updating the standard patterns stored in the memory unit 16. Therefore, in this embodiment, even if an inappropriate standard pattern is registered in the standard pattern memory section 16 due to the influence of noise or malfunction, the switch I8 is switched to the standard pattern creation data memo 9 section 15 side. , by creating a standard pattern for registration using the data stored here.
It becomes possible to re-register from the initial state.

[Effects of the Invention] As explained above, according to the present invention, the number of speakers can be easily increased, and the number of speakers to be recognized can also be easily increased. It is possible to provide a speech recognition device that does not require an increase in the capacity of standard patterns in proportion to the number of standard patterns.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a first embodiment of the present invention, FIG. 2 is a diagram showing the storage state of standard patterns in the standard pattern memory section of the device of the above embodiment, and FIG. 3 is a block diagram showing the configuration of a first embodiment of the present invention. FIG. 4 is a block diagram showing the configuration according to the second embodiment, and FIG. 4 is a block diagram showing the configuration according to the third embodiment of the present invention. 11...Signal converter, 12.18.Switch, 13.
・・Similarity calculation unit, 14 ・Standard pattern creation unit, 1
5...Memory section for standard pattern creation data, 1B...
- Standard pattern memory section, 17...judgment section.

Claims (7)

[Claims] (1) a standard pattern creation unit that creates a standard pattern for registration from input audio; a first memory unit that stores data used when creating the standard pattern for registration in the standard pattern creation unit; a second memory section that stores in advance standard patterns corresponding to various sounds and also stores the standard patterns created by the standard pattern creation section; A speech recognition device comprising: a similarity calculation unit that calculates similarity with a speech input pattern. (2) When voices of the same vocabulary by different speakers are input, the new standard pattern for registration created by the standard pattern creation section is set to the standard pattern of the corresponding vocabulary in the second memory section. 2. The speech recognition device according to claim 1, wherein the speech recognition device is configured to be re-stored in a pre-stored area. (3) A memory unit that stores in advance standard patterns corresponding to various voices, and a standard that creates new standard patterns using input voices and the standard patterns stored in the memory unit and stores them again in the memory unit. A speech recognition device comprising: a pattern creation section; and a similarity calculation section that calculates the similarity between a standard pattern stored in the memory section and a speech input pattern for recognition. (4) The standard pattern creation unit is configured to change the standard pattern stored in the second memory unit and create a new standard pattern when the same vocabulary is input by different speakers. The speech recognition device according to claim 1 or 3. (5) The speech recognition device according to claim 1 or 3, wherein the standard pattern creation section creates a standard pattern for registration from input speech using a dominant speaker equality method. (6) The speech recognition device according to claim 1 or 3, wherein the standard pattern creation section creates a standard pattern for registration from input speech using a dominant speaker identification method. (7) The speech recognition device according to claim 1 or 3, wherein the standard pattern creation section creates a standard pattern for registration from input speech using a smoothing/differential pattern orthogonalization method.