JPH064096A - Voice recognizer - Google Patents

Abstract

(57) [Abstract] [Purpose] The purpose of the speech recognition apparatus is to improve word recognition performance. [Structure] In a speech recognition apparatus that matches an input speech pattern with a standard pattern using the DP method, and uses the standard pattern with the smallest matching distance as a recognition result, the input pattern is divided into phonemes using the matching result, and each phoneme is divided. It is characterized in that the variance of the difference between the continuous time and the standard continuous time is calculated, and this is added to the matching distance to correct the distance. Dividing unit 6
The phoneme is divided into phonemes using the matching result in the time length deviation calculation unit 7
Calculate the variance of the deviation from the standard duration with
Is configured to correct the matching distance. Further, it has a phoneme selection unit 9 for selecting a target phoneme for calculating the time length deviation and a word selection unit 10 for selecting a target word for distance correction.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing a word by matching an input voice pattern with a standard word pattern.

[0002]

2. Description of the Related Art FIG. 4 is a block diagram of a conventional voice recognition apparatus. FIG. 5 shows the collation data used therein.

The spectrum analysis unit 1 divides the input voice into fixed time periods (frames), performs spectrum analysis using FFT or the like for each frame, and holds the analysis result.
As a unit of analysis, a frame length is about 10 milliseconds and a frequency division is about 200 to 5000 Hz. A band is divided into about 20 channels, and the power value of the division range is used as a value. The frequency band is divided by equal division or mel scale division (division according to the sensitivity of the human ear).

The analysis result is, as shown in FIG.
Is an input frame number and j is a frequency division number (channel number), and is represented by A = {a _ij }.

In the phoneme template storage unit 2, FIG.
As shown in (B), a standard speech pattern processed by the same division method as the input is held as a template for each phoneme or for each phoneme unit (hereinafter referred to as a phoneme) corresponding to the phoneme.

Phonemes are divided into about 20 categories such as vowels (A, I, U, E, O) and consonants (K, S, T, N, H, M ...) A plurality of templates of about 10 types are prepared for each category in response to deformation and the like due to the influence of other phonemes.

As shown in FIG. 5B, the template is represented by B = {b _klj } where k is a phoneme category (type) number, l is a template number in each category, and j is a frequency division number. Be done.

[0008] duration storage unit each phoneme in 3, the k as a category number, the minimum length of the duration, the maximum length {s _k, t _k} Mean duration (standard duration) {v _k} and, FIG. 5 (C) and 5 (D), k is a category number and m is a duration (the number of frames), and a duration weight distribution G = {g _km } is stored. These are called standard duration distributions.

As shown in FIG. 5E, the word model storage unit 4 stores a phoneme sequence C = {c _wn } for each word model, where w is a word number and n is a phoneme number in the word model. ing.

The collation unit 5 collates the input voice with each word model, and the one with the highest degree of similarity is regarded as the identification result of the input voice. The problem is that the same word has different durations for each utterance. Therefore, non-linear pattern matching on the time axis is necessary.

Since the input pattern A and the standard pattern C _w are associated on the time axis one by one, DP (Dynamic Programi)
ng) method is used. FIG. 6 is a conceptual diagram of matching by the DP method. Considering the time-series correspondence between the input pattern A and the standard pattern C _w , if the time axis of the input pattern A is expanded or contracted, and if a path that connects the frames at the intersection (lattice point) and finds the shortest connection is found, Good. For that purpose, the cumulative value of the total distance is minimized based on the distance (local distance) between the input frame and the phoneme template. This is done for each word model and the one with the smallest distance is taken as the result.

For each word model, the local distance, that is, the distance d _in between the input frame i and the closest phoneme template of the node n (nth phoneme) of the word model is the category number k of the phoneme of that node. , And the number of phoneme templates is L, (k: category number of the nth phoneme of the word model).

The shape of the DP path becomes that shown in FIG. 6 by controlling the duration. The value of each grid point indicates the shortest cumulative distance to reach it. The cumulative distance D _in to each grid point is However, y: continuous time (number of frames) initial value of cumulative distance D ₀₀ = 0 D _i0 = ∞ (i = 1, I) D _0n = ∞ (n = 1, N).

Considering the cumulative distance at the grid point (i, n), that is, D _in , the first term of the equation (2) is based on the constraint of the duration (s _k to t _k ), and the one before it reaches this point. It indicates that the grid points are limited. That is, the grid point immediately before the path reaching the grid point in FIG. 6 is limited to any of, and the other grid points are excluded.

The second term is the distance converted from the average (standard value) of the duration times into the distance, and the third term is the cumulative distance to the previous grid point. The matching distance with the word model w is a cumulative distance obtained by comparing the final frame of the input and the final phoneme of the word model, and D _w = D _IN (W: word number, I: final frame of the input,
N: final phoneme of word model) The matching result is D = {D _w }.

Of the matching results, the word model with the smallest matching distance is output as the recognition result.

[0017]

The following problems may occur when word matching is performed by the above-mentioned conventional technique. For example, when inputting the voice "Oota", instead of "OOTA",
It may be mistakenly recognized as "OOITA". In this case "O
The matching distance D ₁ (for example, 50) with “OITA (hereinafter word number 1)” is smaller than the matching distance D ₂ (for example, 60) with “OOTA (hereinafter word number 2)”. In this case, "OOITA" is obtained as the recognition result. 7A and 7B show examples of input results and collation results with each word in this case. "OOITA", "O
The duration of each phoneme of "OTA" is "10,10,10,7,23" and "15,15,7,23", respectively.

(Unit: frame 1 frame = 10 ms) Here, the standard duration of each phoneme is (O) (I) (T) (A) v ₅ = 12, v ₂ = 12, v ₈ = 4 , v ₁ = 20
In case of (unit: frame), deviation from standard duration Z _wn = v _k -dur _n Expression 3 (w: word number, n: node number in word model k: category number of n-th phoneme of word model v _k : standard duration of category k dur _n : duration when the word model is the nth phoneme) is (O) Z ₁₁ = 12-10 = 2 (O) Z ₂₁ = 12-15 =- 3 (O) Z ₁₂ = 12-10 = 2 (O) Z ₂₂ = 12-15 = -3 (I) Z ₁₃ = 12-10 = 2 (T) Z ₂₃ = 4-7 = -3 (T) Z ₁₄ = 4-7 = -3 (A) Z ₂₄ = 20-23 = -3 (A) Z ₁₅ = 20-23 = -3 Unit: Frame. This is shown in FIG. 7 (C).

When the input voice is pronounced fast or slow, the deviation from the standard duration is one direction for each phoneme, but when it is collated with another word, the deviation is not. It shows that the direction (and size) varies.

As described above, even if the collation distance is small, the collation result may be incorrect if the variation from the average with respect to the duration is large. It is an object of the present invention to realize a voice recognition device having a high recognition rate by considering the variation from the average regarding the duration.

[0021]

FIG. 1 is a block diagram showing the principle of the present invention. With respect to the conventional speech recognition apparatus, a division unit 6 that divides an input voice pattern into phonemes, a time length deviation calculation unit 7 that calculates a variance of deviation from a standard duration, and a distance correction unit 8 that corrects a matching distance. With.

[0022]

The reason why the conventional speech recognition apparatus obtains an incorrect recognition result is that the matching distance is used for each phoneme without considering the variation in deviation from the standard duration.

In order to solve the above problem, the variance SD _w of the deviation from the standard duration is calculated for each phoneme, (ave Z _w : average of deviation Z _wn from standard continuation time) Add to conventional collation distance D _w as a correction distance.

ND _w = D _w + k SD _w (k: proportional constant) Formula 5 By this, a new matching distance ND _w that more accurately represents the degree of similarity in consideration of the variation in the deviation of the duration is obtained. .

FIG. 3 shows an operation explanatory diagram of the dividing unit 6. Figure 3
(A) DP when the input is associated with "OOITA"
Indicates a path. The minimum value D of the cumulative distance to each grid point
_It indicates in, and the route with the shortest cumulative distance as a whole is indicated by a frame and an arrow.

FIG. 3B shows the value Y _{in of the} duration y when the cumulative distance D _in of the equation 2 becomes the minimum value at each grid point. The dividing unit 6 receives this Y _in from the matching unit 5 and stores it. After this is collated, as shown in FIG. 3 (B), the path is moved backward from the end (word ending) to the start (word beginning) to divide into phonemes. In this way, the phoneme duration {dur _n } when corresponding to each node of the word model is obtained.

The time length shift calculation unit 7 calculates the difference Z _wn between the duration dur _n of the divided phonemes and the standard duration v _k obtained from the duration storage unit 3 by the equation 3, and further the equation 4 Then, the deviation variance SD _w is obtained.

The distance correction unit 8 calculates the deviation SD by the formula 5
_The corrected matching distance ND _w is calculated by adding _w, and the word closest to the distance is used as the recognition result. With this configuration, a more accurate recognition result can be obtained.

When obtaining the deviation variance SD _w , the processing time can be shortened in practice by limiting the target phonemes or target words.

[0030]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 2 is a block diagram of an embodiment of the present invention. Components having the same functions as those in FIGS. 1 and 4 are denoted by the same reference numerals.

In FIG. 2, reference numeral 1 denotes a spectrum analysis unit, which includes an AD converter, an FFT operation circuit, and the like, and a storage unit for storing the analysis result. Reference numeral 2 is a phoneme template storage unit, 3 is a duration storage unit, and 4 is a word model storage unit, which is placed in the storage unit of the EWS (engineering workstation) together with the storage unit of the analysis result. Reference numeral 5 is a collation unit, 6 is a division unit, 7 is a time length deviation calculation unit, and 8 is a distance correction unit.
It consists of a processor and software.

The operation procedure is shown below. The spectrum analysis unit 1 analyzes the spectrum, and the result A =
Hold {a _ij }. The frame length is 10 as an analysis unit
20 milliseconds band with a frequency range of 200-5000Hz
It is divided into about channels, and the power value in the divided range is used as a value. The phoneme template B = {b stored in the phoneme template storage unit 2, the duration storage unit 3, and the word model storage unit 4
_klj }, word model C = { _cwn }, minimum duration,
Using the maximum length {s _k , t _k } and the weight G = {g _km },
The collation unit 5 performs DP collation to obtain D 1 = {D _w }. There are about 20 categories, about 10 templates, and about 1000 word models. Up to this point, the process is the same as the conventional technique. Next, the dividing unit 6 divides the input speech into each phoneme by using the matching result as shown in FIG. The variance is calculated, and the distance correction unit 8 corrects the matching distance using the result calculated by the time shift calculation unit 7.

For example, in the case of the above "Ota", "O
Since the deviation relating to "OITA" varies, the distance is increased, and the distance relating to "OOTA" remains unchanged because there is no variation in deviation. Specifically, each matching distance is ND ₁ = D ₁ + k SD ₁ = 50 + 2 × 6 = 62 ND ₂ = D ₂ + k SD ₂ = 60 + 2 × 0 = 60 (aveZ: average of deviation Z from standard duration), and there is nothing smaller than ND ₂ (in this case, the matching distance D ₃ with other word models is much larger). , The matching result can be correct as "Ota".

In this embodiment, the phoneme selection unit 9 and the word selection unit
I added 10. Of course, only one of them is enough. The phoneme selection unit 9 limits the phonemes for which the deviation from the standard should be calculated. This is because vowels and the like are significantly different from consonants having a short duration. For example, if it is limited to only vowels, the processing time becomes short.

The word selection unit 10 limits the words for which the distance correction is performed. This is because it is not necessary to perform distance correction on all word models, and it is sufficient to perform distance correction on a higher one of the matching distances D _w (those that are likely to give a correct recognition result). Processing time may be short if correction is performed only for the top few words.

[0036] For example, the distance D _w is about the words in the following range value distance D _w is less difference in the distance D _w of the 1-position for the word to top n in the order of about words following frequency values rank up one Distance correction may be performed for words whose difference in distance D _{w from} the word is a threshold value or less.

[0037]

As described above in detail, according to the present invention, by performing the post-processing of the conventional collation method by reflecting the deviation of the duration time on the collation distance, more precise collation becomes possible and the recognition rate is improved. It is possible to realize a high voice recognition device.

[Brief description of drawings]

FIG. 1 is a block diagram of the principle of the present invention.

FIG. 2 is a block configuration diagram of an embodiment of the present invention.

FIG. 3 is an operation explanatory diagram of a dividing unit.

FIG. 4 is a block diagram of a conventional voice recognition device.

FIG. 5 is a diagram showing collation data.

FIG. 6 is a conceptual diagram of matching by the DP method.

FIG. 7 is a diagram showing an example of a matching result.

[Explanation of symbols]

 1 spectrum analysis unit 2 phoneme template storage unit 3 duration storage unit 4 word model storage unit 5 collation unit 6 division unit 7 time difference calculation unit 8 distance correction unit 9 phoneme selection unit 10 word selection unit

Claims (6)

[Claims] 1. A spectrum analysis unit (1) for performing spectrum analysis of input speech and storing it as characteristic time series data.
And a phoneme template storage unit (2) that stores feature data of a phoneme or a phoneme unit that is similar to a phoneme, and a duration storage unit (3) that stores a standard duration distribution of a phoneme or a phoneme unit that is similar to a phoneme, a word or A word model storage unit (4) that stores a model of a speech unit corresponding to a word, and a matching unit that matches the spectrum analysis result of the input speech with the word model by controlling the duration using a phoneme template and duration distribution ( In the speech recognition device having 5), a dividing unit (6) that divides the input speech into phonemes or phoneme units corresponding to phonemes using the matching result, and duration and standard continuation of the divided phonemes or phoneme units corresponding to phonemes. There is a time length deviation calculation unit (7) for calculating a time difference and a distance correction unit (8) for correcting the matching distance using the calculation result. Speech recognition apparatus characterized by. 2. The speech recognition apparatus according to claim 1, further comprising a phoneme selection unit (9) for specifying a phoneme or a phoneme unit corresponding to the phoneme for which a difference between the duration and the standard duration is to be calculated. 3. The voice according to claim 1, further comprising a word selection unit (10) for setting a word or a voice unit corresponding to the word whose distance is to be corrected to have a matching distance equal to or less than a predetermined threshold value. Recognition device. 4. The word selection unit (10) for setting the order of the matching result within a predetermined order for a word or a speech unit corresponding to the word for which distance correction is to be performed. Speech recognizer. 5. A word selection in which a word to be subjected to distance correction or a voice unit corresponding to the word is a word whose difference in matching distance from the matching distance of a word ranked first in a matching result is equal to or less than a predetermined threshold value. Claim 1 characterized in that it has a part (10).
Voice recognition device. 6. The word or the speech unit equivalent to the word for which the distance correction is performed is up to a difference in the matching distance from the matching distance of the word having a higher rank in the matching result is equal to or less than a predetermined threshold value. The speech recognition apparatus according to claim 3, further comprising a word selection unit (10).