JPS63236095A - Voice recognition - Google Patents

Abstract

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

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a word speech recognition method in a word speech recognition device characterized by using a word dictionary with phonetic notation, and in particular to a word speech recognition method based on the assumption of words. This relates to a top-down speech recognition method.

[Conventional technology]

Conventional speech recognition methods can be roughly divided into the following two types.

(1) Recognition method based on words.

(2) A recognition method that uses units smaller than words, such as phonemes.

The first method has the feature parameters of the entire word as a standard pattern, and performs pattern matching with the input speech using DP, achieving a high recognition rate in word recognition of ~ several hundred words. ing. However, this method is time-consuming because all the words to be recognized must be registered by vocalization. Since words are used as standard patterns, the capacity of the word dictionary is large, making it unsuitable for recognizing human language of 1,000 words or more.

On the other hand, the second method has a standard pattern for each phoneme, so
Since word dictionaries can be registered in phonological notation, there is no need for registration and the capacity is small, so compared to the first method,
It can be said that it is suitable for people who speak a lot of words.

In the second method, it is common to first perform segmentation to determine the range in which each phoneme exists, and then perform phoneme recognition. On the other hand, a top-down recognition method can be considered. In this method, a word to be recognized is assumed, and a standard pattern of phonemes included in this word is used to perform matching with the feature time series of input speech, and a score is calculated. Furthermore, recognition is performed by determining the score for the entire word based on the score for each phoneme. In this way, since only the standard pattern of phonemes necessary for the word is used, compared to methods that perform segmentation first, additions and omissions of phonemes occur and recognition accuracy increases.

As an example of this, for example, "1-op-down acoustic processing in a speech recognition system" by Sugiyama et al.
(Acoustical Society of Japan Speech Study Group Materials 582-62).

Also, as an example of applying DP matching to top-down processing as a method to reduce the amount of processing, Moriai et al.'s "Automatic labeling of isolated uttered words using a phoneme discrimination filter" (Acoustical Society of Japan Speech Study Group Materials 585-54) There is. Although this research by Moriai et al. did not directly deal with speech recognition, this method can be applied as is to top-down speech recognition.

Top-down speech recognition using this DP matching will be explained below. In this method, the similarity calculation formula is determined using DP matching, for example, by the following incremental formula.

g (i, O) = g (0, D −-OO(i≠o
, j≠0) ・・・・・・・・・・・・・・・・・・(4) However, I: Length of phoneme symbol string, J: Number of frames of input speech V: Phoneme symbol string (■l + v2 + ”' V i
+ ”' V l )X; Characteristic time series of input speech (X
,,X2. ...Xi, ...XJ) g (i, D: cumulative similarity of DP, s (v, x): similarity of v and X A (1+j): similarity of X' and Xj, this g (
For example, the window restriction when calculating i and D is i/2≦j≦(i-1-1)/2+J+1...
...(5) etc. are used.

Word recognition is performed based on the similarity obtained as a result.

[Problem to be solved by the invention] The above window restriction is
This is effective not only in reducing the amount of calculation by simply limiting the calculation area, but also in preventing extreme matching. However, in the case of words that are uttered in isolation, the duration of each phoneme is not constant within the word (generally speaking, the duration of each phoneme is longer at the end than at the beginning. "Phonological duration control for speech synthesis" (Transactions of Institute of Electronics and Communication Engineers '84)
/7 Vol. J67-A No. 7) for details.

For this reason, the linear window restriction shown in equation (5) may not be able to deal with variations in the duration of each phoneme depending on its position within a word, and especially when the window width is narrowed to avoid extreme matching, There was a problem in that the input speech could not be accurately mapped to each phoneme.

[Means (and effects) for solving the problem] The present invention has a table for each number of moras that records the temporal position of each phoneme in isolated utterance words, and inputs the values obtained from the table. The above objective is achieved by adapting to the input word by expanding or contracting it in accordance with the duration of the word, and further by setting a weighted DP window using the adapted value and the weighting function.

[Effect]

As a result, by using the above-mentioned DP window, even in sounds where the duration of each phoneme within a word is different, such as an isolated utterance word, the expansion and contraction of the duration can be kept within the window, and the similarity using DP Calculations are made more precisely.

〔Example〕

FIG. 1-1 is an explanatory diagram of one embodiment of the present invention.

1 is an acoustic analysis unit that performs frequency analysis of input utterances, 2 is a feature extraction unit that extracts features necessary for phoneme identification, and 3 is a preliminary selection that narrows down candidate words for the word dictionary based on the feature extraction results. 4 is a word dictionary registered by phonetic symbol strings, 5 is a mora number counting unit that counts the number of moras of the selected word, and 6 is a mora that stores the start and end times of mora corresponding to each mora number. The start/end time table 7 shows the start/end time of the selected mora, the weight function, the audio start time Ts (ms), and the audio end time Te(
8 is a weighting function table that stores weighting functions;
9 is a phoneme standard pattern table that stores standard patterns corresponding to each phoneme, and 10 is a window restriction obtained from 7 that matches the feature time series obtained by extracting features from input speech with the phoneme standard pattern. This is a word recognition unit that uses DP. FIGS. 1-2 are block diagrams showing specific configurations for carrying out the recognition method of the present invention. In Fig. 1-2, 11 is an input section for inputting audio, 12 is a disk for storing various data, and 13 is a control section for controlling this device (system), which includes a control program as shown in Fig. 1-3. It includes a ROM that stores . 14 is a RAM for storing various data in each section shown in FIG. 1-1.

15 is an output section. In addition, each part shown in FIG. 1-1 may have a CPU, RAM, and ROM, respectively.

The present invention will now be described in detail with reference to FIGS. 1-3.

In the above configuration, the sound is frequency-analyzed by one acoustic analysis unit (Sl). At this time, the audio start time ts and the audio end time te are also determined (S2). A second feature extraction unit extracts features necessary for phoneme recognition from the speech spectrum obtained by frequency analysis (S3). Using the global features of the feature time series obtained here, the preliminary selection section 3 selects a plurality of words from the word dictionary 4 whose features are considered to be similar to the input speech (S4). Since the word dictionary No. 4 is registered by phoneme symbol strings, the number of moras N of the selected word can be determined by counting the number of vowels, pellic sounds/long vowels, and consonants. In this way, the mora number N of the word selected in the mora number counting section of 5
is determined (S5), and a value indicating the temporal position of the start and end of each mora corresponding to this number of moras is selected from the six mora start/end time tables (S6). The phoneme estimation interval calculation unit 7 calculates the estimated existence interval of each phoneme based on the value selected in 6 (S7).

The operations 6 and 7 will be further explained in detail using FIGS. 2 to 5.

The mora start and end time table No. 6 stores the mora start and end times corresponding to the words of each mora number. The reason why a plurality of times are set depending on the number of moras of a word is that the state of expansion and contraction of the duration of a phoneme differs depending on the number of moras. This selection from the plurality is performed using the count result of the mora number counting section 5.

FIG. 2 shows, as an example, words corresponding to trimolar words. The horizontal axis of this graph represents time and the vertical axis represents mora. t s (m i ), t e (m i )
respectively indicate the start and end times of the i-th mora. 7
ts adapted to the input speech by linearly expanding and contracting the start and end of the horizontal axis in FIG.
(mi) /le Determine the position of (mi). ts (
After determining the positions of mi) and te(mi), each mora section is divided into phoneme units. The division method is shown in Figure 2 (1
) to (3). For moras, plosives, and consonants that contain only vowels and no consonants, the motor interval was treated as the phonological interval. Furthermore, moras containing consonants other than plosives and obsessives are divided into 1/2 to form consonant and vowel sections, respectively. In addition, the mora containing a persistent consonant was divided into 1/3 to form a consonant section, a persistent consonant section, and a vowel section, respectively.

The weighting functions stored in the 8 weighting function tables were applied to each phoneme interval determined in this manner to form a weighted DP window. FIG. 3 shows an example of a weighting function. The horizontal axis of the figure represents time and the vertical axis represents weight. By linearly expanding and contracting the horizontal axis and aligning the tps and tpe shown in the figure with the start and end times of each phoneme, it is converted into a weighting function suitable for each phoneme interval. This state is shown in FIG.

Next, at 88 in Fig. 1-3, the word recognition unit of IO uses the phoneme estimation interval obtained by the phoneme interval calculation unit of 7 as a window restriction, and uses the characteristic time series of the input speech obtained in 2. DP matching is performed using patterns included in the word selected as a result of preliminary selection among the phoneme standard patterns registered in the phoneme standard pattern table of and 9.

In DP matching, calculations are performed by weighting scores according to the weighting function described above. The score of each candidate word was calculated, and the one with the highest score was selected as the recognized word (S9, 5IO)
.

In this embodiment, a word preliminary selection section is provided, so that candidate words for matching are narrowed down and recognition can be performed at high speed even with a human vocabulary.

Furthermore, it can be used as an automatic speech labeling system by following the above-mentioned DP path and attaching the start and end symbols of each phoneme to the beginning and end of that phoneme.

〔Effect of the invention〕

As explained above, according to the present invention, in the case of isolated uttered words, a non-linear window restriction is used in DP calculation, taking into consideration the property that the phoneme is longer at the end of the word than at the beginning of the word. Even if the window width is narrowed to avoid an increase in Since the score decreases as the phoneme interval deviates from the estimated phoneme interval, a route with higher estimation reliability is selected, resulting in a more accurate DP value.

As a result, the effect of improving the word recognition rate can be obtained.

Fig. 1 is an explanatory diagram of the present invention, Figs. 1-2 are block diagrams of the configuration to which the present invention is applied, Figs. 1-3 are control flowcharts of the present invention, and Fig. 2 is a diagram of the phoneme in the case of a trimolar word. FIG. 3 is a diagram showing an example of a weighting function; FIG. 4 is an explanatory diagram of a weighted window of DP for the word (/ky a no N /).

1... Acoustic analysis section, 2... Feature extraction section, 3... Preliminary selection section, 4... Word dictionary, 5... Mora number counting section, 6... Mora start/end time table , 7... Phoneme estimation interval calculation unit, 8... Weighting function table, 9... Phoneme standard pattern table, 10... Word recognition unit.

Ji-hui-ken language Voice recording “Unagiju’s buO…Figure 7” Fu7-Z diagram Ichiri Takeko 70 - Noon Yahi Lower 1 - child drawing 11L%M (m51

Claims (2)

[Claims] (1) Having a standard pattern for each phoneme expressed as a time series of feature vectors and a word dictionary expressed by phoneme symbols, and the degree of similarity between the speech input pattern expressed as a time series of feature vectors and registered words. In calculating,
Assuming a candidate word corresponding to the input speech, the existence of each phoneme is determined from the speech rate calculated from the duration of the input word and the number of moras of the candidate word in order to recognize the word using the standard pattern of phonemes included in the word. A speech recognition method that estimates the interval, determines the window-limited interval, and then calculates the degree of similarity using dynamic programming (hereinafter abbreviated as DP). (2) The window used in DP is expressed by a weighting function in the direction of the time axis, and is designed to prevent extreme deviation from the estimated interval as it moves away from the center of the estimated existence interval in time. The speech recognition method according to claim 1, characterized in that the value of the weighting coefficient is reduced.