JP2601448B2 - Voice recognition method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

TECHNICAL FIELD The present invention relates to a speech recognition method and an apparatus thereof. 2. Description of the Related Art A so-called BTSP (Binary Time-Spectru) is used in which a speech is binarized to obtain a characteristic pattern, and the input pattern and the dictionary pattern obtained by the binarization are linearly matched and recognized.
The present applicant has already proposed various types of speech recognition using the “m Pattern” method. However, in the BTSP method, since the voice is subjected to the binary processing, information of energy and power representing the volume of the voice may be lost, resulting in erroneous recognition. For example, it is difficult to distinguish the difference between / P /, which is a plosive sound in which a consonant part rises rapidly, and / K /, which rises relatively slowly. Therefore, it is conceivable that the power information of the voice is binarized by a normal method and held together with a binary TSP (BTSP) pattern. In this case, however, since the calculation method of the BTSP and the power unit is different, the power unit A dedicated calculation unit for obtaining similarity is required, which complicates the apparatus and loses the high-speed calculation performance of the BTSP method. SUMMARY OF THE INVENTION The present invention has been made in view of the above-described circumstances, and in particular, it has been proposed to add speech energy or power information to a recognition system using a binary pattern represented by binary values of "1" and "0". However, it is an object of the present invention to provide a recognition method and a device capable of high-accuracy and high-speed recognition. Configuration According to the present invention, in order to achieve the above object, (1) a feature amount of a voice is extracted and stored as a standard pattern;
In a speech recognition method in which a recognition result is determined by collating with a speech pattern of an unknown input speech, the energy or power shape of speech is expressed as a two-dimensional pattern having two axes orthogonal to time and power magnitude. 2 corresponding to the energy or power of the sound at time
A position on the dimensional pattern or its vicinity and other parts are represented by "1" and "0" to create a two-dimensional patterned voice pattern, and one or more patterns created by the same procedure are overlapped. The sum of the standard patterns is used as a standard pattern, the unknown input voice is also converted into a two-dimensional pattern in the same manner, and the similarity is determined by superimposing the standard pattern on the unknown pattern.
(2) In a voice recognition device that extracts a feature amount of a voice and holds it as a standard pattern, and determines a recognition result by comparing the voice pattern with a voice pattern of an unknown input voice.
A power detector for detecting the energy or power of the voice, and a voice section, and expressing the power shape of only the portion related to the voice section as a two-dimensional pattern having two axes orthogonal to time and power magnitude; "1" indicates a position on or near the two-dimensional pattern corresponding to the energy or power of the sound at the time, and other parts.
And a switching unit for switching between the time of standard pattern creation and the time of recognition, and a plurality of utterance patterns created by the binarizing unit when the standard pattern is created. A standard pattern creating unit for creating a standard pattern that is superimposed on and added to the two-dimensional pattern, and a two-dimensional pattern created by the binarizing unit for an unknown input speech when the unknown speech is recognized. A superimposition unit for collating the standard pattern created by the pattern creation unit, a similarity determination unit for determining the similarity of the pattern superimposed by the superimposition unit, and a maximum similarity determined by the similarity determination unit And a recognition result output unit that uses the standard pattern of (1) as a recognition result. (3) In (2), after the voice section is cut out A power detection unit and a feature amount conversion unit are provided side by side, and the two-dimensional pattern is used together with a two-dimensional pattern created by another feature amount to determine a similarity between patterns of each type. And (4) determining a recognition result by determining the final similarity by applying the other similarity to the power detection unit after the voice section is cut out in (2). A power pattern matching unit for judging similarity by comparing the two-dimensional standard pattern with the unknown input two-dimensional power pattern; and a similarity measure based on the result of the power pattern matching unit. And a similarity determining unit that determines whether to calculate the similarity. When the two-dimensional pattern is used in combination with a two-dimensional pattern created by another feature amount, when similarity is determined between patterns of each type, One Similarity score when a particular condition is satisfied, or is obtained by said determining a recognition result by also calculated seen other similarity when not satisfied. Hereinafter, a description will be given based on examples of the present invention. FIG. 1 is an electric block diagram for explaining an embodiment of the present invention, in which 1 is a microphone, 2 is a power detector, 3 is a voice section detector, 4 is a binarizer, 5 is a register, 6 is an addition unit, 7 is a standard pattern, 8 is a superimposition unit, 9 is a similarity determination unit, and 10 is a recognition result output unit.
The power of the voice input from the microphone 1 is detected, the voice section is obtained, and only the portion related to the voice section is binarized and represented by a binary pattern of "1" and "0". The power detection may be obtained by, for example, detecting the envelope of the amplitude of the audio waveform, and the detection of the audio section is obtained as a portion where the power obtained by the power detection unit 2 is equal to or more than a certain value. or,
In the binarization unit, the portion indicating the shape of the audio power is represented by "1" and the others are represented by "0" as in the binarization information shown in FIG. 2B from the power signal shown in FIG. 2A. In this example, the magnitude of the power is quantized in five stages, and it is easy to associate the waveforms from (B) to (A). When creating a standard pattern, switch S

[Picture] Flip to the side, and one voice, for example / Pa /
Three times. First, the first utterance pattern is put into the register 5, superimposed and added with the second pattern, and put into the register 5 again. Next, the third pattern and the contents of the giresta 5 are added and registered as a standard pattern. That is, the standard pattern (E) is created by adding the patterns (B), (C), and (D) in FIG. 2, and the speech is registered after repeating this for each voice to be registered. In the case of recognition, an unknown voice is input from the microphone 1 and passes through the same process as when the standard pattern is created, and is "1",
The pattern binarized to “0” is checked against some standard patterns created earlier. In the collation, the pattern binarized into “1” and “0” and one of the standard patterns are superimposed to calculate the similarity. The binary "1" and "0" patterns of the unknown voice at this time are shown in FIG. 2 (B).
If both have similar waveforms, the element of "1" in the pattern of (B) overlaps with the element showing a large value of (E) by overlapping with the standard pattern (E). become. Therefore, as the similarity, the product of the elements corresponding to each other by superposition of the two may be taken and defined as the sum thereof. The similarity between all the registered standard patterns and the unknown “1” and “0” patterns is obtained, and the one with the highest similarity is output as a recognition result. By this, binarization processing (1,0 processing)
Then, the power information is added and the similarity can be calculated. However, it is difficult to recognize speech only based on power information. FIG. 3 is an electric block diagram showing another embodiment in which the above-mentioned disadvantages are solved. In this embodiment, the pattern created in the above embodiment and the pattern created by other feature amounts are used together. The similarity between the patterns is calculated, and one similarity is made to act on the other similarity to obtain the final similarity. In this case, another method used in combination with the conventional technique is described as 2 The method using the TSP of values was chosen. This is a result of the binarization processing performed together with the binary TSP power pattern, and the same operation is possible. However, it is needless to say that other methods may be used in combination. In FIG. 3, after a voice section is cut out by the voice section detection unit 3, power detection is performed by the power detection unit 2, and the same signal is subjected to feature value conversion by the feature value conversion unit 11. In this embodiment, a spectrum is suitable for the feature amount. The power and the shape of the spectrum are binarized by the binarization unit 4 into “1” and “0”. In the binarized pattern of “1” and “0”, the subsequent calculation is easier if the spectral pattern and the power pattern are combined into one pattern. Binarization unit 4
The example of the pattern in FIG. 4 is as shown in FIG. 4, where F is a normal BTSP and G is a pattern corresponding to the pattern in FIG. 2 (B).
The similarity may be calculated in the same procedure as in the example of FIG. 1 to derive the result. In this case, if the similarity determination unit 9 considers that the size of the pattern has increased, there is no part that changes the procedure, and the result is obtained based on the similarity of the sum of the two patterns. Thereby, the accuracy is dramatically improved as compared with the example of FIG. In this case, weights can be placed on either the power or the spectrum pattern and the others can be used as auxiliary means. FIG. 5 is an electric block diagram for explaining an embodiment based on the above viewpoint. According to this embodiment, when similarity is obtained, one of the similarities satisfies a specific condition or Only when the condition is not satisfied, the other similarity can be calculated to determine the recognition result. In this embodiment, similar to the embodiment shown in FIG. 3, "1" and "0" patterns as shown in FIG. 4 in which the spectrum and the power are combined are created and registered several times. At the time of recognition, the power part of the unknown input power generated by the binarizing unit 4 is compared with the power pattern matching unit 12 to see the similarity. If the similarity is significantly different, the determination unit 13 determines that the similarity of the spectrum part is not calculated, and the process proceeds to the collation with the next standard pattern. If the determination unit 13 determines to calculate the similarity between the spectral patterns, the similarity between the patterns is calculated as in FIG. In this case, the similarity may be calculated with or without the power part. In this example, the power of the entire pattern is compared, but it goes without saying that this may be performed for each frame in the pattern instead of for one entire audio pattern. Effects As is clear from the above description, according to the present invention, power information is added to the spectral pattern binarized to “1” and “0”, and the accuracy of speech recognition can be improved.

[Brief description of the drawings]

FIG. 1 is an electric block diagram for explaining an embodiment of the present invention, FIG. 2 is a diagram showing a binarization pattern for explaining the operation of the present invention, and FIG. FIG. 4 is an electric block diagram for explaining another embodiment of the present invention.
FIG. 5 is a diagram showing an example of a binarization pattern, and FIG. 5 is an electrical block diagram showing another embodiment of the present invention. 1 ... microphone, 2 ... power detector, 3 ... voice section detector,
4 binarization unit, 5 register, 6 addition unit, 7 standard pattern, 8 superposition unit, 9 similarity determination unit, 10 recognition result output unit, 11 feature amount conversion unit, 12 ... Power pattern collating unit, 13 ... judgment unit.

 ──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-60-19199 (JP, A) JP-A-59-222900 (JP, A) JP-A-59-205680 (JP, A) JP-A-59-205680 186073 (JP, A) Proceedings of the Acoustical Society of Japan (October 1983) 3-1-8, p. 195-196

Claims (4)

(57) [Claims] 1. A speech recognition method for extracting a feature amount of a speech, storing the feature amount as a standard pattern, and determining a recognition result by comparing the feature with a speech pattern of an unknown input speech. The time and the magnitude of the power are expressed as a two-dimensional pattern having two axes orthogonal to each other, and the position on or near the two-dimensional pattern corresponding to the energy or power of the sound at each time and the other parts are represented by “ A two-dimensional pattern-formed voice pattern is represented by “1” and “0”, and one or more patterns created by the same procedure are superimposed and added as a standard pattern. It is characterized in that it is converted into a dimensional pattern, the similarity is determined by superimposing it on a standard pattern, and the maximum similar standard pattern is used as the recognition result. Speech recognition method. 2. A speech recognition apparatus for extracting a feature amount of a speech, storing the extracted feature amount as a standard pattern, and determining a recognition result by collating with a speech pattern of an unknown input speech. And a power detector that calculates a voice section and calculates the power shape of only the portion related to the voice section by orthogonalizing the time and the magnitude of the power.
The two or more axes are expressed as two-dimensional patterns, and the position on or near the two-dimensional pattern corresponding to the energy or power of the sound at each time, and other parts are expressed by “1” and “0”. A binarizing unit for creating a dimensional pattern, a switching unit for switching between a standard pattern creation time and a recognition time, and a plurality of utterance patterns superimposed on the two-dimensional pattern created by the binarizing unit when the standard pattern is created. A standard pattern creating unit for creating the added standard pattern, and a two-dimensional pattern created by the binarizing unit for the unknown input speech at the time of recognition when unknown speech is input, and a standard created by the standard pattern creating unit. A superimposition unit for collating the pattern, a similarity determination unit for determining the similarity of the pattern superimposed by the superimposition unit, and the similarity Speech recognition apparatus characterized by comprising a recognition result output unit for a recognition result a maximum similarity of the reference pattern determined by the tough. 3. A power detecting section and a feature quantity converting section are arranged side by side after the speech section is cut out, and the two-dimensional pattern is used together with a two-dimensional pattern created by another feature quantity. A similarity between patterns of a kind is obtained, and a similarity is determined by applying one of the similarities to the other to determine a final similarity, thereby determining a recognition result. A speech recognition device according to the item. 4. After the speech section is cut out, a power detection section and a feature quantity conversion section are provided side by side, and the two-dimensional standard pattern is compared with an unknown input two-dimensional power pattern to determine similarity. A two-dimensional pattern having a power pattern matching unit and a similarity determining unit for determining whether to calculate similarity based on the result of the power pattern matching unit, When determining the similarity between patterns of each type, when one similarity satisfies a specific condition or only when it does not, the other similarity is calculated and the recognition result is determined. The speech recognition device according to claim 2, wherein the speech recognition device is characterized in that: