JP2721341B2 - Voice recognition method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method, and more particularly to a speech recognition method for recognizing speech based on the acoustic characteristics of input speech. 2. Description of the Related Art Conventionally, this type of apparatus has a disadvantage that all items registered in a dictionary are to be recognized, and if all entries are searched, it takes a long time to perform an input voice recognition process. Therefore, when targeting a large vocabulary, the processing speed of the current computer is:
There is a problem that real-time property, which is an important element of voice, cannot be guaranteed. Therefore, we first roughly extract features from the input speech, perform rough matching calculations with words in the dictionary, and for those that give good results, perform matching with detailed features again. Although the speed was increased by the reduction, even with this method, it was necessary to always perform the matching calculation once for every word, and furthermore, there was a disadvantage that it was difficult to incorporate into a speech recognition device having phoneme recognition means. . Japanese Patent Application Laid-Open No. 61-13 / 1986 discloses a technique in which words are classified according to a certain feature in advance so that matching is not performed for all words, and recognition candidates are preliminarily selected based on the feature.
No. 7198, JP-A-61-149997, JP-A-61-67899 and JP-A-60-217399. [Problems to be Solved by the Invention] However, since the preliminary selection in the above-mentioned conventional example is intended only to speed up speech recognition by a large classification (rough classification), each word has a feature. , But only its commonality has priority. For example, in Japanese Patent Application Laid-Open No. 61-137198, only words predicted based on a vowel / pronunciation sequence are preliminarily selected and their syllables are compared and compared. , A candidate corresponding to the inserted or replaced vowel / prompt string is added to the result of the preliminary selection. Also, JP-A-61-
In 149997, a vowel symbol string candidate is extracted from the vowel large classification and the consonant large classification, but the extraction result covers possible replacement examples from each vowel and consonant. In such a simple merging of the selection candidates, since it is not considered how the voice pattern of each word easily changes or how it is extracted differently, the range of the merging is limited. If it is enlarged, words that do not originally need to be candidates for recognition matching increase, which hinders speeding up. On the other hand, if the range of merging is limited, correct words are excluded at the time of preliminary selection, and The danger of leaking. The present invention eliminates the above-mentioned drawbacks of the prior art, speeds up recognition by performing preliminary selection of words based on the acoustic characteristics of the input voice, and how the voice pattern of each word is changed. By performing a preliminary selection in consideration of the tendency to change or how it is extracted differently, even if the features used for the preliminary selection are extracted differently or there is a personal difference in the input speech. However, the present invention provides a speech recognition method that enables high-speed and high-recognition speech recognition. [Means for Solving the Problems] In order to solve the problems, a voice recognition method according to the present invention represents a voice dictionary storing word voice patterns for matching and an acoustic feature of the input word voice patterns. A word index and an index dictionary that stores a pointer indicating a word voice pattern of the voice dictionary associated with the word index are prepared, and in the index dictionary,
Pointers indicating the relevant word voice parameters of the voice dictionary are stored in a plurality of word indexes respectively representing a plurality of voice features that can be extracted from one word voice pattern, and a voice feature is extracted from the input word voice pattern. Matching a word voice pattern of the voice dictionary indicated by the pointer with the input word voice pattern according to a pointer stored in the index dictionary corresponding to the word index representing the extracted acoustic feature. Then, the input word voice pattern is recognized. Embodiment FIG. 1 shows a configuration diagram of a speech recognition apparatus according to a first embodiment. In FIG. 1, reference numeral 1 denotes an audio input unit, which comprises a microphone 2 and an A / D converter 3, and converts input audio into digital. Reference numeral 4 denotes a sound processing unit for extracting frequency conversion and acoustic parameters. Reference numeral 5 denotes a preliminary selection unit, which comprises a preliminary selection feature extraction unit 6, a preliminary selection index creation unit 7, and a word search unit 8. The preliminary selection unit 5 uses the audio feature amount output from the audio processing unit 4 for preliminary selection. Extract features and create an index from those 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 obtains a pointer to a word in the dictionary 10 included in the index from the index dictionary 9, The word in the dictionary 10 pointed to by the pointer to the word is output to the candidate word memory 11 as a candidate word.
Preliminary selection is performed by the method described above. 12 is a phoneme recognition unit,
The output of the acoustic processing unit 4 is used as input to recognize phonemes and output a sequence of phonemes. A word matching unit 13 matches a candidate word in the candidate word memory 11 with a phoneme sequence output from the phoneme recognition unit 12, and outputs a word recognition result. FIG. 2 shows an example of acoustic features used for preliminary selection and their encoding. The unvoiced friction in the figure (about "s"
“C”, “h”), voiced friction (“z”), voiced vowels, semi-vowels, rhymes (“z”, “b”, “d”, “g”, “r”), Unvoiced sounds (“s”, “c”, “h”, “p”, “t”, “k”), silence and rhyme (“m”, “n”, “N” of the repellent sound) ") Is an acoustic feature that is relatively easy to extract, although the appearance of the feature differs depending on the phonemes before and after the phoneme. In addition, a symbol is given to each of these acoustic features by adding a time duration. FIG. 3A shows a method of creating an index based on the notation.
In this example, "I""watasi" to "wa" are long voiced sounds [V], [q] for short silence between "wa" and "ta", and short voiced sounds for "ta". [V] and “si” are represented as a long unvoiced friction [S] and a short voiced sound [v], and the word “watasi” is referred to as a preliminary selection index [VqvSv] (hereinafter INDEX). ) To create. At this time, when "watasi" is uttered, the feature extraction may be different. Considering a likely event at this time, for example, a short interval between “s” and de-voicing of “i” of “si”, [V
Create INDEX such as [qvsv] or [VqvS]. “Baketu”
An example of a “bucket” is also shown in FIG. FIG. 4 shows the configuration of the index dictionary 9 and the configuration of the dictionary 10 using INDEX shown in FIG. 3 (b). FIG. 5 shows an operation flowchart of the voice recognition apparatus of the first embodiment. In step S51, "I" from the voice input unit 1 in this example
When the voice input is performed, the audio processing is performed by the audio processing unit 4 in step S52. The result of the acoustic processing is directed to feature extraction by the feature extraction unit 6 in step S53 and phoneme recognition by the phoneme recognition unit 12 in step S59. In step S53,
Extract features such as unvoiced friction, voiced friction, voiced, unvoiced, silence, Buzz, nasal, and duration of the input voice,
In step S54, an INDEX is created according to the table shown in FIG. Here, an INDEX called “VqvSv” is usually created for the input voice of “I (watasi)” as shown in FIG. 3A, but the voiceless fricative of “s” is short and “ Vqvsv ”will be described as an example.
In this case, in step S56, the index dictionary 9 is searched based on INDEX "Vqvsv". In steps S56-S57, in FIG. 4, INDX "Vqvsv" → pointer 10 → "watasi", INDEX "V
Preliminarily select recognition candidates in the order of “qvsv” → pointer 20 → “baketu”,..., and store candidate words “I”, “bucket”,. On the other hand, the result of the phoneme recognition in step S59 and the recognition candidates stored in the candidate word memory 11 are subjected to matching judgment by the word matching unit 13 in step S58, and the recognition result is obtained in step S60. Is output. Since the word matching in the system shown in FIG. 1 is based on phoneme symbols, a phoneme recognition unit 12 is required. However, if the word expression in the dictionary 10 is represented by acoustic level parameters and the word matching unit 13 performs acoustic parameter level matching, the phoneme recognition unit 12 becomes unnecessary.
In addition, the present invention is not limited to the example shown in FIG. 2 and may be used to symbolize, for example, burstiness (may be divided into voiced and unvoiced) and the magnitude of energy (power) of voice. FIG. 6 shows a configuration diagram of the speech recognition apparatus of the second embodiment. In the figure, reference numeral 61 denotes a sound input unit which is constituted by a microphone 62 and an A / D converter 63, and the input voice is converted to digital. Reference numeral 64 denotes a sound processing unit for extracting frequency conversion and acoustic parameters. 65 is a preliminary selection unit, which is 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 feature bit sequence comparison unit that determines whether or not a word extracted from the dictionary 69 requires more detailed matching calculation as a candidate word.
A feature bit sequence is extracted from the acoustic feature amount output from the audio processing unit 64, and a feature bit sequence is created from the extracted feature. Feature bit sequence comparison unit
Reference numeral 68 compares a feature bit sequence output from the feature bit sequence creation unit 67 with a feature bit sequence assigned to a word in the dictionary 69, and a "1" is added to the feature bit sequence on the input voice side. If all "1" s are set in the feature bit sequence on the dictionary 69 side for all the words, the word having the feature bit sequence is determined as a candidate word and sent to the candidate word memory 70. Reference numeral 71 denotes a phoneme recognition unit that performs phoneme recognition using the output of the acoustic processing unit 64 as an input, and outputs a phoneme sequence. Reference numeral 72 denotes a word matching unit that matches a word on the candidate word memory 70 with a phoneme sequence output from the phoneme recognition unit 71 and outputs a word recognition result. FIG. 7 shows an example of the configuration method of the dictionary 69 in FIG. This example is an example of the word “bento”. The characteristic bit sequence for the phonetic symbol "beNtou" of this "lunch box" is H
eader (it does not need to be 16 bits) and in this example, a bit sequence represented by 16 bits (16 bits for each phoneme symbol). Here, one 16-bit (corresponding to a phoneme) is defined as SEG. Each SEG is represented by 16 bits,
The 16 bits have a meaning as shown in FIG. 7 for each of 0 to 15 bits. For example, the 0th bit drops off and a 1 here indicates that the SEG may drop off. The first and second bits are longer or shorter in the duration of the feature. The reason why “1” is set in both SEG1s is that they may become longer or shorter. In other words, we are thinking about the possibility that Buzz may or may not occur for "b" because it is the beginning of a word. Similarly, considering the dropout of the previous SEG and the influence of the SEG before and after, the rupturability and rhinosylation of SEG2 are incorporated. Each SEG can be created if the phoneme sequence is automatically determined based on the acoustic characteristics of each phoneme and the relationship between before and after. As the information of Header,
In consideration of the dropout of SEG, the maximum and minimum SEG numbers that can occur as a whole and the pointer to the word are entered. A dictionary structure having a characteristic 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 “beNtou” is generated. Unlike the dictionary of FIG. 7, the representation method in FIG. 8 is the same except for considering the possibility of insertion. However, HE
No pointer to the word in the ADER part is required. Each SEG
In the decision section, "1" bit is set for the obtained feature. FIG. 9 shows a SEG sequence that can be represented by the represented SEG sequence since insertion and omission are assumed on the input side and the dictionary side, respectively. Using these SEG sequences, it is determined whether or not to perform detailed matching. FIG. 10 is a diagram for explaining the operation of the characteristic bit sequence comparing section 68 which is a portion for determining whether or not to perform detailed matching.
First, the SEG sequence of the input voice is put into the input voice side register 104, and the S
The EG sequence is retrieved from the dictionary 69. The SEG sequences are sequentially stored in the memories 102 and 103. The bits in the memories 102 and 103 are compared by a comparison circuit 105, and an output 106 of "1" and "0" is obtained. FIG. 11 (a) shows the comparison circuit 105 and its operation explanatory diagram.
(B). If the determination at 107 is “0”, the next word or the next SEG sequence is extracted from the input voice side or the dictionary side at 108 or 109, and a repeating process is performed. “1” at 107
If it is the judgment of 110, CHECK whether it is the last SEG at 110. If it is not the last SEG, the same processing is performed on the next SEG at 113. By repeating the above, for words for which the output of the comparison circuit 105 is all “1” until the final SEG, 111 is used as a candidate word.
At 12 the word is sent to the word matching section 72 for detailed matching. It should be noted that the method of using acoustic features for dictionary contents can be applied to a system without the phoneme recognition unit 71 in FIG. As described above, by providing a preliminary selection unit and an index dictionary in the speech recognition apparatus, a word having a possibility of having the acoustic feature of the input speech is preselected, and the number of word matching is reduced by setting the word candidate as a word candidate. , Enabling faster recognition processing time. In addition, by using the detailed matching determination unit in the speech recognition device, the amount of waste can be greatly reduced, and the recognition processing time can be reduced. [Effects of the Invention] According to the present invention, a word is preliminarily selected in an index dictionary based on the acoustic characteristics of an input voice to increase the speed of recognition, and how the voice pattern of each word changes. Preliminary selection using an index dictionary that considers how easy it is to extract or how it is extracted differently, even if the features used for preselection are extracted differently, It is possible to provide a speech recognition method that enables high-speed and high-recognition speech recognition even if there is a significant difference. That is, in the index dictionary for performing the preliminary selection,
Since pointers are stored in consideration of erroneous extraction in acoustic feature extraction, useless matching is minimized, the possibility of correct words leaking at the time of preliminary selection is drastically reduced, and a high recognition rate can be obtained at high speed Voice recognition becomes possible.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of a speech recognition device according to a first embodiment, FIG. 2 is a diagram showing an example of symbolizing acoustic features according to the first embodiment, FIG. 3 (a) , (B) is a diagram showing a method of creating an index of the first embodiment, FIG. 4 is a configuration diagram of an index dictionary and a dictionary of the first embodiment, and FIG. 5 is a speech recognition device of the first embodiment. FIG. 6 is a block diagram of the speech recognition apparatus of the second embodiment, FIG. 7 is a diagram showing the notation of a word dictionary of the second embodiment, and FIG. 8 is a second embodiment. FIG. 9 is a diagram showing an example of the feature notation of the input speech of FIG. 9, FIG. 9 is a diagram showing the SEG sequence in the second embodiment, FIG. (A), (b) is an explanatory view of the circuit A of FIG. In the figure, 1 ... voice input unit, 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 unit, 9 ... index dictionary, 10 ... dictionary, 11 ... memory for candidate words, 12 ... phoneme recognition unit, 13 ... word matching unit, 14 ... recognition result output unit, 61 ... voice input Department,
62: microphone, 63: A / D converter, 64: sound processor, 65: preliminary selector, 66: feature extractor, 67: feature bit sequence generator, 68: feature bit sequence comparator, 69: dictionary, 70: candidate word memory, 71: phoneme recognition unit, 72: word matching unit,
73 ... Recognition result output unit.

Claims (1)

(57) [Claims] A speech dictionary storing a word speech pattern for matching, a word index representing the acoustic feature of the input word speech pattern, and a pointer indicating the word speech pattern of the speech dictionary associated with the word index are stored. In the index dictionary, pointers indicating the word voice parameters of the voice dictionary are stored in a plurality of word indexes respectively representing a plurality of acoustic features that can be extracted from one word voice pattern. Extracting an acoustic feature from the input word speech pattern, and according to a pointer stored in the index dictionary corresponding to a word index representing the extracted acoustic feature, a word in the speech dictionary indicated by the pointer The voice pattern is matched with the input word voice pattern, and Speech recognition method and recognizes the spoken word pattern.