JPS6385697A - Voice recognition equipment - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device, and particularly to a speech recognition device that reduces recognition time.

[Prior Art] Conventionally, devices of this type have the disadvantage that all items registered in a dictionary are subject to recognition, and that it takes time to recognize input speech when searching all the entries. For this reason, when targeting dog language food, there is a problem that current computer processing speeds cannot guarantee real-time performance, which is an important element of speech.

Therefore, we first roughly extract features from the input speech, perform rough matching calculations with words in the dictionary, and if good results are obtained, we perform matching again using detailed features. Speed reduction was achieved by reducing the number of words, but even with this method, matching calculations had to be performed once for every word, and there was also the drawback that it was difficult to incorporate into speech recognition devices that had phoneme recognition means. .

[Problems to be Solved by the Invention] The present invention eliminates the above-mentioned conventional drawbacks, performs preliminary word selection based on the acoustic characteristics of speech input from a dictionary, and simply reduces the number of word candidates. To provide a speech recognition device that reduces the number of word matching processes and shortens speech recognition time.

[Means for Solving the Problem] As a means for solving this problem, the speech recognition device of the present invention includes a feature extracting means for extracting acoustic features from input speech, and a feature extracting means for extracting acoustic features from input speech. and a recognition means that recognizes a recognition result from the recognition candidates preselected by the preliminary selection means.

[Operation] In the speech recognition apparatus of the present invention, the recognition means recognizes a recognition result from recognition candidates preselected in advance by the preliminary selection means based on the acoustic features extracted by the feature extraction means.

[Example] FIG. 1 shows a configuration diagram of a speech recognition device according to a first embodiment.

In the figure, reference numeral 1 denotes an audio input section, which is composed of a microphone 2 and an A/D converter 3, and converts input audio into digital data. 4 is a sound processing section which performs frequency conversion and extracts sound parameters. Reference numeral 5 denotes a 1,000m selection unit, which consists of a feature extraction unit 6 for preliminary selection, an index creation unit 7 for preliminary selection, and a word search unit 8.
Extract features for preliminary selection and create an index from the features. The word search unit 8 searches the index dictionary 9 for an index that matches the index output from the index creation unit 7, and sends a pointer to the word in the dictionary 10 that the index has to the index dictionary 9.
The word in the dictionary 10 pointed to by the pointer to that word is output to the candidate word memory 11 as a candidate word. A preliminary selection is made in the manner described above. 12 is a phoneme recognition unit which receives the output of the acoustic processing unit 4, performs phoneme recognition, and outputs a series of phonemes. 13 is the word matching section,
Candidate words on candidate word memory 11 and phoneme recognition unit 1
Matching is performed with the phoneme sequence output from step 2, and the word recognition result is output.

FIG. 2 shows an example of odd habits used for preliminary selection and their symbolization. The unvoiced friction in the diagram (roughly expressed as a phoneme is “S”,
“, corresponds to “h”), voiced friction C (“Z”), voiced vowel, semi-vowel, nasal rhyme (“z”,
"b", d", "", 'r"), unvoiced sounds ("S", "
, “h”, “p”.

“t”, “k”), silent/nasal @ (“N” in “m”, “n” sound… phoneme equivalent to “person” after Japan), etc., are pronounced depending on the phoneme before and after the phoneme. Although the appearance behavior of features is different, it is an acoustic feature that is relatively easy to extract. Furthermore, we added the temporal duration to these acoustic features and assigned symbols to each of them.

FIG. 3(a) shows a method of creating an index based on notation. In this example, #wa# from "watashi" is a long voiced sound [V], [q] is a short silence between "wa" and "ta", and a short voiced sound is "ta". [V]
, 'si'' is expressed as a long voiceless fricative [S] and a short voiced sound [V], and for the word "watasi", a preliminary selection index of [VgvSv] (hereinafter IN
DEX). Next, when "watasi" is uttered, the feature extraction may be different. Events that are likely to occur at this time, such as short intervals between “S”
Considering things like devoicing the “i” in “i”, we set [Vqvsv
] or [Create an INDEX with VqvS.

"baketu""An example of packet j is also shown in FIG. 3(b).

FIG. 4 shows the structure of the index dictionary 9 and the structure of the dictionary 10 using the INDEX shown in FIG. 3(b).

FIG. 5 shows an operation flowchart of the acoustic recognition apparatus of the first embodiment.

In step S51, from the voice human power section 1, in this example, "I"
If there is voice power, the sound processing unit 4 performs sound processing in step S52, and the result of the sound processing is processed in step S5
The process proceeds to feature extraction by the feature extraction unit 6 in step S59 and phoneme recognition by the phoneme recognition unit 12 in step S59. Step S53
INDEX is created in step S55 from the result of feature extraction. In this example, rVgvSv] is created. In step S56, the index dictionary 9 is searched based on INDEX rVgvSv], and in steps 556-57, the fourth
In the figure, INDEX rVgvSv → pointer 10
→"watasi"-I NDEXrVgvsv] → Pointer 20 → "baketu" → I NDEX
r B v g v s v ] −*−・−, the recognition candidates are preliminarily selected in order, and the candidate words “me”, “bucket”,
... is stored in the candidate car word memory 11.

On the other hand, the results of phoneme recognition in step S59 and the recognition candidates stored in the candidate word memory 11 are used in step S55.
In step S60, the word matching section 13 determines whether or not the word is matched, and in step S60, the recognition result, in this example, "Washi" is output.

Note that since the word matching in the system of FIG. 1 is based on phoneme symbols, the phoneme recognition unit 12 is required. However, if the word expressions in the dictionary 10 are represented by acoustic level parameters and the word matching unit 13 performs matching at the acoustic parameter level, the phoneme recognition unit 12 becomes unnecessary. In addition to the examples shown in Figure 2, for example, plosiveness (which can be divided into voiced and unvoiced) and vocal energy (power)
The size of the image may be symbolized.

FIG. 6 shows a configuration diagram of the voice recognition bag chamber of the second embodiment.

61 in the figure is the audio input section, which includes micro 2 and A/D converter 6.
3, the input audio is converted into digital data.

Reference numeral 64 is a sound processing unit that performs frequency conversion and extracts sound parameters. Reference numeral 65 denotes a preliminary selection unit, which includes a feature extraction unit 66 for obtaining a feature bit sequence to be determined as a word candidate, a feature bit sequence creation unit 67, and a dictionary 69.
The feature bit sequence comparison unit 68 determines whether a word extracted from the word requires more detailed matching calculation as a candidate word, and the acoustic processing unit 64
A feature for a feature bit sequence is extracted from the acoustic features output from the system, and a feature bit sequence is created from the extracted feature. The feature bit sequence comparison unit 68 is the feature bit sequence creation unit 67
The feature bit series that is the output of is compared with the feature bit series that is attached to the words in the dictionary 69, and for all the feature bit series that are "1" on the input voice side,
If all 1'' are set in the feature bit series on the dictionary 69 side, the word having that feature bit series is determined as a candidate word and sent to the candidate word message processor 70. 71 is a phoneme recognition unit that performs acoustic processing. Phoneme recognition is performed using the output of the unit 64 as input, and a phoneme sequence is output.A word matching unit 72 performs matching between the word in the candidate word memory 70 and the phoneme sequence output from the phoneme recognition unit 71, and then outputs a phoneme sequence. Output recognition results.

FIG. 7 shows an example of how the dictionary 69 of FIG. 6 is constructed. This example is about the word "lunch box." The feature bit series for this phoneme symbol "beNtou" for "lunch box" is a header (not necessarily 16 bits) and a bit series represented by 16 bits in this example (16 bits for each phoneme symbol). It is expressed as. Here, one 16 bit (corresponding to a phoneme) is assumed to be an SEG. Each SEG is represented by 16 bits, and each of the 16 bits has a meaning as shown in FIG. 7, including 0 pits and 15 bits.

For example, the 0th bit is dropped, and a value of 1 here indicates the possibility that the SEG will be dropped. 1st bit 1
The second bit indicates whether the duration of the feature is long or short. In 5EGI, "1" is set in both fields because there is a possibility that the length may become longer or shorter.

In other words, since it is the beginning of a word, we are considering the possibility that 'bH may or may not have Buzz. Similarly the previous SEG
Considering the dropout of 5EG2 and the shadow 1 of SEG before and after, we incorporate the rupture and rhinism of 5EG2. Each SEG can be created if the phoneme sequence is clearly determined from the acoustic characteristics and context of each phoneme. Header
The information includes the maximum number of SEGs, the minimum number of SEGs that can occur as a whole, and a pointer to a word, considering the separation of SEGs from feudal domains. As described above, a dictionary structure having a feature bit sequence as shown in FIG. 7 is created.

FIG. 8 shows an example of a feature bit sequence created from input speech. This example is an example of a feature bit sequence expected when "b e N t u" is generated.

Unlike the dictionary shown in FIG. 7, the method of representation is the same in FIG. 8 except for considering the insertion possibility bit. However, HE
Pointers to words in the ADER section are not required. In each SEG determination interval, a "1'° bit is set for the obtained feature.

FIG. 9 shows an SEG sequence that can be expressed by the expressed SEG sequence since insertions and omissions are assumed on both the input and dictionary sides. It is determined whether detailed matching is to be performed using the stiff SEG sequence.

FIG. 10 is an explanatory diagram of the operation of the feature bit sequence comparison unit 68, which is the part that determines whether or not to perform detailed matching. First, the input audio register 104 contains the input audio S.
The EG sequence is input into the dictionary side register 101, and the SEG sequence whose length matches that of the input side is extracted from the dictionary 69. Each SEG series is sequentially stored in memories 102 and 103. Comparison circuit 1 compares bits in memories 102 and 103
05 and an output 106 of "1" and "0" is output. FIGS. 11(a) and 11(b) are diagrams illustrating the comparison circuit 105 and its operation.
). If it is O'' in the judgment of 107, then 108
Alternatively, in step 109, the next word or the next SEG sequence is read out to the input voice side or dictionary side, and repeated processing is performed. 1
If the judgment is “1” at 07, check whether it is the final SEG at 110.
HECK. If it is not the last SEG, proceed to the next SEG with 113.
Perform the same process on EG. Repeat the above and
Words for which the output of the comparison circuit 105 is all 1'' up to the final SEG are sent as candidate words in 111 to the word matching unit 72 in 112 for detailed matching.

In addition, in the method of using acoustic features in the dictionary contents,
It is also possible to apply the present invention to a system without the phoneme recognition section 71 in FIG.

As explained above, by providing a preliminary selection unit and an index dictionary in the speech recognition device, words that may have the acoustic characteristics of the input speech are preselected and used as word candidates, thereby reducing the number of word matching operations. , which made it possible to speed up the recognition processing time.

Furthermore, by using a detailed matching determination section in the speech recognition device, many unnecessary measurements can be greatly reduced, and the recognition processing time can be shortened.

[Effects of the Invention] According to the present invention, words are preliminary selected based on the acoustic characteristics of the speech input from a dictionary, and the number of word candidates is reduced, thereby reducing the number of word matching processes and shortening the speech recognition time. A voice recognition device can be provided.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the speech recognition device of the first embodiment, FIG. 2 is a diagram showing an example of symbolizing acoustic features of the first embodiment, and FIGS. 3(a) and (b) are the first embodiment. FIG. 4 is a configuration diagram of the index dictionary and the dictionary in the first embodiment. FIG. 5 is an operation flowchart of the speech recognition device in the first embodiment. FIG. 6 is a block diagram of the speech recognition device of the second embodiment, FIG. 7 is a diagram showing the notation of the vulgar dictionary of the second embodiment, and FIG. 8 is an example of notation of characteristics of input speech in the second embodiment. FIG. 9 is a diagram showing the SEG sequence in the embodiment of S2, FIG. 10 is an explanatory diagram of circuit A in FIG. 10. FIG. In the figure, 1... audio input section, 2... microphone, 3...
A/D conversion unit, 4... Sound processing unit, 5... Preliminary selection unit, 6... Feature extraction unit, 7... Index creation unit, 8...
・Word search section, 9... Index dictionary, 10... Dictionary, 1
1... Candidate word memory, 12... Phoneme recognition unit, 1
3... Word matching unit, 14... Recognition result output unit, 61... Audio input unit, 62... Microphone, 63...
- A/D conversion unit, 64... Sound 9 processing unit, 65... Preliminary selection unit, 66... Feature extraction unit, 67... Feature bit sequence creation unit, 68 Feature bit sequence comparison unit, 69 ...
Dictionary, 70... Candidate word memory, 71... Phoneme recognition section, 72... Word matching section, 73... Recognition result output section. Patent applicant Canon Co., Ltd. Figure 2 Figure 3 (0) Figure 3 (b) Figure 4 16 Dobuto 01 2345678 9A...
EF Figure 8 Figure 9 Factor 11 (a) Figure 11 (b)

Claims (4)

[Claims] (1) In a speech recognition device that performs speech recognition by comparing input speech with a pre-registered dictionary, there is a feature extraction means for extracting acoustic features from the input speech, and features extracted by the feature extraction means. 1. A speech recognition device comprising: a preliminary selection means for preselecting recognition candidates based on characteristics of the selected sound; and a recognition means for recognizing a recognition result from the recognition candidates preselected by the preliminary selection means. (2) The speech recognition device according to claim 1, wherein the feature extraction means extracts features from the input speech as a string of predetermined standard sounds in units of words. (3) The speech recognition device according to claim 1, wherein the feature extraction means extracts the features by decomposing the input speech into predetermined standard sounds. (4) The recognition means recognizes the recognition result based on a comparison between the recognition candidates preselected by the preliminary selection means and the phoneme recognition of the input speech. recognition device.