JP2886879B2 - Voice recognition method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech recognition method, and more particularly to a consonant vowel syllable (hereinafter C) of an unspecified speaker.
The present invention relates to an effective speech recognition method for recognizing speech composed of V syllables.

[Prior Art] Conventionally, this type of speech recognition method is an extraction target.
Frequency analysis of voices composed of CV syllables, such as voiced plosives, and extraction of the first to third formant frequencies from spectral peaks obtained by linear prediction analysis (hereinafter referred to as LPC analysis). It is already known to track the change and identify the voiced plosive from its slope (IEICE Technical Report SP88-14).

[Problem to be Solved] By the way, when the above-mentioned method is adopted in recognizing a voice composed of CV syllables of an unspecified speaker, a temporal change of the second formant frequency is directly followed. Therefore, not only does the formant frequency itself differ greatly among men, women, children, and the like, but also frequency fluctuations such as intonation are directly susceptible, and erroneous recognition of voice is likely to occur.

The present invention has been made in view of the above problems, and provides a novel speech recognition method capable of reducing erroneous recognition in recognizing speech composed of CV syllables of an unspecified speaker. It is.

[Means for Solving the Problems] The present inventor has analyzed the problems described above as follows, and has come up with the present invention.

More specifically, in the conventional speech recognition method, erroneous recognition occurs because the second formant frequency itself is used in the feature space for recognizing speech, and the time variation of the formant frequency, which is an important feature of consonants, occurs. Including the crossover
This is probably because the speech composed of CV syllables was not normalized.

Therefore, in order to normalize the speech composed of the CV syllables, it is important to select a feature parameter of the speech that does not depend on a specific speaker. Considering that it corresponds to the frequency axis (hereinafter referred to as auditory correspondence scale) space that matches the auditory characteristics of
The present inventors have found out a speech characteristic parameter obtained by converting a formant frequency that is not directly dependent on a specific speaker into a relative value, and have come up with the present invention.

In other words, the present invention provides a method for recognizing a voice composed of CV syllables, wherein, among the formant frequencies obtained by analyzing the frequency of the voice signal, the auditory correspondence between the adjacent first and second formant frequencies and the second and third formant frequencies The difference information in the scale is extracted as a feature parameter, the vowel component from the convergence point of the difference information in the auditory correspondence scale of the adjacent formant frequency is identified, and the consonant component based on the rising position and the temporal behavior of the difference information. Is identified.

In such technical means, the CV syllable to be recognized may be voiced or unvoiced as long as the first to third formants occur. The auditory scale includes a mel scale or a logarithmic scale approximating the mel scale.

In addition, the order of the step of discriminating the vowel component and the consonant component may be any order, but since the temporal behavior of the characteristic parameters related to the consonant component varies depending on the vowel component, from the viewpoint of easy identification, It is preferable to identify a consonant component after identifying a vowel component first.

Further, the means for implementing the characteristic parameter extracting step and the vowel component / consonant component identifying step can be appropriately changed in design as long as they can exhibit the necessary functions in each step.

[Operation] According to the technical means as described above, the characteristic parameter of the voice is the difference information of the first formant frequency and the second formant frequency, and the second formant frequency and the third formant frequency on the auditory correspondence scale. The feature parameters of speech are normalized in accordance with the auditory space without directly depending on the formant frequency, and are distributed with temporal behavior in the feature space of the auditory correspondence scale. At this time, the temporal behavior of the feature parameters is
It rises at a predetermined location in the feature space, moves with time, and converges at the predetermined location.

The convergence position of the feature parameter corresponds to the vowel component of the CV syllable, and the variation in the feature space falls within a relatively narrow range.

On the other hand, the position of the characteristic parameter on the way to the rising position and the convergence position corresponds to a consonant component of the CV syllable, and the variation in the characteristic space falls within a relatively narrow range.

Hereinafter, the present invention will be described in detail based on an embodiment shown in the accompanying drawings.

FIG. 1 is a block diagram showing an embodiment of an apparatus for realizing a speech recognition method according to the present invention.

In the figure, reference numeral 1 denotes a microphone for converting a sound into an electric signal, and 2 denotes an amplifier and a low-pass filter for cutting off unnecessary high frequency components after amplifying the sound signal input from the microphone 1 and adjusting the amplitude level of the sound signal. 3 is an AD converter that converts an analog signal from the signal adjustment circuit 2 into a digital signal, 4 is a frequency branching unit that analyzes frequency by linear prediction analysis, and 5 is a frequency analysis unit. A parameter extraction unit 6 extracts a feature parameter from the obtained formant frequency and records the change over time, and a speech identification unit 6 identifies a speech based on information of the parameter extraction unit.

In this embodiment, the parameter extracting section 5 performs a control operation according to a flowchart shown in FIG.

Specifically, the parameter extraction unit 5 first extracts the first to third formant frequencies f1 to f3 based on the result obtained by the frequency analysis unit 4 (step 1), and then extracts the formant frequency The mutual distance log (f1 / f2) (hereinafter referred to as Lf12) and log (f3 / f2) (hereinafter referred to as Lf23) on a logarithmic scale which is an approximation of the mel scale of f1 to f3 as first and second feature parameters. After the calculation, it is stored in the memory (steps 2 and 3).

After that, the parameter extraction unit 5
It is determined whether or not Lf12 and Lf23 are the first ones (step 4). If it is the first time, the process returns to step 1, and if not, it is determined whether or not the characteristic parameters Lf12 and Lf23 have changed. Is determined (step 5).

If the characteristic parameters Lf12 and Lf23 are changing, the parameter extracting unit 5 clears the count number described later (Step 8), and repeats the operation in Steps 1 and later in a predetermined cycle, and repeats the operation in a predetermined cycle. The time change of the first and second characteristic parameters Lf12 and Lf23 are sequentially stored in the memory area of.

On the other hand, if the characteristic parameters Lf12 and Lf23 have not changed, the parameter extraction unit 5 counts up by “1” (step 6), and determines whether or not the count has reached a predetermined specified number m. After the discrimination (step 7), when the number reaches the specified number m, the vowel component is determined to have been input, and the parameter extraction process is terminated.
To return to.

Further, the voice identification unit 6 performs a control operation according to a flowchart shown in FIG.

Specifically, the voice identification unit 6 first searches for data near the end in the memory written by the parameter extraction unit 5, and determines the feature parameters Lf12 (n) and Lf2 located at the convergence point.
3 (n) is extracted (step 1), and a vowel component is determined by judging which vowel region the feature parameters Lf12 (n) and Lf23 (n) at this convergence point are included in in advance. Step 2) Then, the initial data in the memory written by the parameter extracting unit 5 is searched, and the characteristic parameters Lf12 (1) and Lf23 at the rising position are searched.
(1) One or a plurality of consonant candidates are selected by judging which consonant candidate region is included in advance (step 3). If there is only one consonant candidate, this consonant candidate is recognized. (Steps 4 and 5)
If there are a plurality of consonant candidates, the parameter extracting unit 5 searches the written intermediate area data in the memory and obtains a change pattern of each of the characteristic parameters Lf12 (i) and Lf23 (i) (steps 4 and 6). ), A consonant component to be recognized is determined by comparing the change pattern with predetermined change pattern data (step 7).

Then, the divided area of each vowel used in this embodiment is
As shown in FIG. 4, for example, each vowel "a, i, u, e, o" by 20 unspecified speakers (male: MALE, female: FEMALE) for each gender
The values of the first and second feature parameters Lf12, Lf23 are respectively obtained, plotted in the voice feature space having the first and second feature parameters Lf12, Lf23 as coordinate axes, respectively, based on each vowel distribution plotted. Vowel range Sa, Si, Su, Se, So
Is divided by predetermined partition lines 1 to l4, and a partition line 1 necessary for specifying each vowel region
Or any of the l4 data is stored in the memory in advance. In FIG. 4, MALE indicates ● (black), F
EMALF is ○ (open), “a” is ○, “i” is △, “u” is □, “e” is ▽, and “o” is ◇.

Also, the area of the candidate consonant used in this embodiment is:
As shown in Fig. 5, the CVs of 20 male and female unspecified speakers
Speech composed of syllables (in FIG. 5, for example, "b
a "" ba "" ga ") are listed at the rising position of the first and second feature parameters Lf12 and Lf23, respectively, in the voice feature space with the first and second feature parameters as coordinate axes. Each of the consonant candidates is plotted, and one or a plurality of consonant candidates are demarcated by a predetermined partitioning line l within a range that can be demarcated based on each plotted voice distribution. In FIG. 5, “ba” is indicated by “●”, “da” is indicated by “○”, “ga” is indicated by “■”, and “ga” is indicated by “■” in FIG. And “da, ga” are separated.

Further, the change pattern data of the consonant candidates used in this embodiment are the first and second characteristic parameters after the rise of the consonant which is also the subject of 10 male and female unspecified speakers.
After obtaining the values of Lf12 and Lf23, respectively, and plotting them in the voice feature space using the first and second feature parameters as coordinate axes, a common change pattern of each consonant is changed based on the group of change patterns of each consonant plotted. The change pattern data corresponding to each consonant is stored in the memory in advance.

Next, a specific speech recognition operation using the speech recognition method according to this embodiment will be described.

In this embodiment, of the voice composed of CV adjustment,
As an example, the case where voiced plosives “ba”, “da”, and “ga” are identified, the audio signal is AD-converted at a sampling frequency of 16.6 KHz and a 12-bit length, and the first to the second obtained as a result of LPC analysis. From the three formant frequencies, the temporal behavior of the first and second feature parameters Lf12 and Lf23 is obtained.

In this case, FIGS. 6 and 7 are graphs showing the temporal behavior of the first and second feature parameters Lf12 and Lf23 with respect to the input audio signals "ba", "da" and "ga". Is a graph showing the temporal behavior of the first and second feature parameters Lf12 and Lf23 on a feature space using both as coordinate axes. In addition, in FIGS. 6 to 8, "ba" is
“Da” is ○ and “ga” is Δ.

Now, assuming that the voice signal is "ba", the voice discriminating unit 6 first determines that the vowel is "Sa" in FIG. 4 from the characteristic parameters Lf12 (n) and Lf23 (n) at the convergence point in steps 1 and 2. "
It is determined that the vowel is included in the area, and based on this, it is determined that the vowel component of the recognition target voice is “a”. Then
The voice discriminating unit 6 selects “b” as a consonant candidate based on steps 3 to 5 and determines that there is one consonant candidate, and then recognizes this consonant candidate as a recognition target consonant component “b”. Is determined. Thus, the voice identification unit 6 identifies that the voice to be recognized is "ba".

Assuming that the voice signal is "da", the voice recognition unit 6 first determines that the vowel is "S" in FIG. 4 from the characteristic parameters Lf12 (n) and Lf23 (n) at the convergence point in steps 1 and 2.
It is determined that the vowel component is included in the “a” region, and based on this, the vowel component of the speech to be recognized is determined to be “a”. After that, when the voice discriminating unit 6 determines in step 4 that there are a plurality of consonant candidates, in step 6, the change pattern of the consonant component of the voice to be recognized is changed to the eighth. After extracting that it has a relatively steep slope as shown by arrow A in the figure,
In the comparison step of step 7, the consonant component of the speech to be recognized is determined to be "d". Thereby, the voice identification unit 6 identifies that the voice to be recognized is "da".

Further, assuming that the audio signal is “ga”, the audio identification unit 6 basically goes through the same process as the above-described identification process of the audio signal “da”. In this case, the audio identification unit 6 In step 6, after extracting that the change pattern of the consonant component of the speech to be recognized has an extremely gentle slope as shown by the arrow B in FIG. 8, the comparison process in step 7 is performed. The consonant component of the speech to be recognized is “g”
Is determined.

Further, in this embodiment, the logarithmic difference information of the formant frequency is used as the first and second feature parameters at the rising position of the voice composed of CV syllables (“ba, da, ga” in this embodiment). As a comparative example, the first and second characteristic parameters include first and second formant frequencies f1 and f2 or second and third formant frequencies f2 and f3.
When it was used directly, the results as shown in FIG. 9 or FIG. 10 were obtained.

At this time, according to the distribution diagram in FIG. 9, it is understood that it is difficult to clearly divide the target consonant candidate area, and according to the distribution diagram in FIG. Although the candidate area can be partitioned by a partition line l ', since the partition line l' is not parallel to each coordinate axis,
The method of drawing the partition line l 'is more complicated than the determination in the Lf12-Lf23 space, and in any case, the superiority of this embodiment is supported.

Note that the parameter extraction unit 5 according to this embodiment includes:
All of the temporal behavior data of the characteristic parameters Lf12 and Lf23 are stored in the memory, but from the viewpoint of suppressing the memory capacity, all data that does not change corresponding to the vowel components are individually stored. Without storing
It is preferable to store only one set of data corresponding to a vowel component.

[Effects of the Invention] As described above, according to the speech recognition method of the present invention, the inter-distance on the auditory correspondence scale of the formant frequency is used as a feature parameter, and the temporal behavior in this parameter space is tracked. As a result, voices composed of CV syllables can be effectively normalized in a state corresponding to hearing, and the effect is directly affected by the difference in formant frequency itself of unspecified speakers and frequency fluctuation such as intonation The situation can be effectively prevented, and erroneous recognition of voices composed of CV syllables can be reduced.

Furthermore, according to the speech recognition method of the second aspect, since the consonant component is identified after the vowel component is identified first, the consonant component having a different temporal behavior for each vowel component is identified first. In comparison, the process of specifying the voice composed of CV syllables can be simplified, and the voice composed of CV syllables can be recognized more efficiently.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of an apparatus for realizing a speech recognition method according to the present invention, FIG. 2 is a flowchart showing a control operation process of a formant extraction unit used in the embodiment, and FIG. FIG. 4 is a flowchart showing a control operation process of a voice identification unit used in the embodiment, FIG. 4 is a graph plotting a logarithmic difference of a formant frequency of each vowel by an unspecified speaker on a logarithmic scale, and FIG. A graph in which a logarithmic difference of a formant frequency with respect to a rising position of a part of syllable voices is plotted on a logarithmic scale. FIG. 6 is a graph showing a temporal behavior of a first characteristic parameter of a specific voice including CV syllables. FIG. 7 is a graph showing the temporal behavior of the second feature parameter of a specific voice composed of CV syllables, and FIG. 8 is a feature composed of a logarithmic scale of the specific voice composed of CV syllables Graph showing the temporal behavior between,
FIG. 9 is a graph in which formant frequencies f1 and f2 are plotted with respect to a rising position of a part of voices composed of CV syllables by an unspecified speaker, and FIG. 10 is a part of rising voices of CV syllables by an unspecified speaker. 5 is a graph in which formant frequencies f2 and f3 are plotted against positions. [Description of Signs] 1... Microphone 2... Signal adjustment circuit 3... AD converter 4... Frequency analysis unit 5... Parameter extraction unit 6.

Continuation of the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 9/02 301 G10L 9/06 JICST file (JOIS)

Claims (2)

(57) [Claims] The present invention relates to a method of recognizing a speech comprising a consonant and a vowel syllable, and, among formant frequencies obtained by frequency analysis of a speech signal, hearing of adjacent first and second formant frequencies and second and third formant frequencies. The difference information in the corresponding scale is extracted as a feature parameter, and the difference information in the adjacent corresponding auditory scale of the formant frequency is identified as the vowel component from the convergence point, and the consonant component based on the rising position and the temporal behavior of the difference information. A voice recognition method characterized in that a voice recognition method is performed. 2. The speech recognition method according to claim 1, wherein a vowel component is first identified, and then a consonant component is identified.