JPH0564800B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Field of Industrial Application> The present invention relates to a speech recognition device that recognizes input speech such as Japanese in syllable units and outputs the recognized speech to an external device.

<Prior art> In conventional speech recognition devices, in order to extract syllable sections of syllables, which are recognition units, from input speech, a speech spectrum of a certain section (hereinafter referred to as
The syllable boundaries are detected using changes in the syllable (simply called spectrum).

<Problems to be solved by the invention> However, in the above-mentioned conventional speech recognition device,
Changes in the input speech spectrum include a mixture of rapidly changing speech and gently changing speech, and it is difficult to accurately detect syllable boundaries by following both of them, and often There is a problem that misunderstandings occur.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a syllable recognition device that can accurately detect syllable boundaries using information from the spectrum, regardless of the degree of change in the spectrum.

<Means for Solving the Problems> In order to achieve the above object, the speech recognition device of the present invention extracts syllable intervals from input speech, and stores feature patterns of the extracted syllables in a memory in advance. In a speech recognition device that recognizes input speech in units of syllables by calculating the similarity with the standard feature pattern that has been extracted, the above-mentioned spectral information is A stable point extraction unit that extracts a stable point sequentially calculates the similarity between the spectral information at the extracted stable point and the spectral information after the stable point, and sets the calculated similarity to a predetermined value. The present invention is characterized in that it includes a syllable boundary detection section that compares and detects the syllable boundary of the next syllable interval to be extracted.

<Operation> When speech is input, the stable point extraction section extracts stable points on the spectral information using changes in spectral information in the input speech following the extracted syllable section. Furthermore, the syllable boundary detection unit sequentially calculates the similarity between the spectral information at the extracted stable point and the spectral information after the stable point, and calculates the obtained similarity value by a predetermined value. By comparing with the value, the phonetic boundary of the next syllable interval to be extracted is detected, and the syllable interval is extracted according to the detected syllable boundary. Therefore, it is possible to accurately extract meaningful syllable sections regardless of the degree of change in the spectral information.

<Examples> The present invention will be described in detail below with reference to illustrated examples.

In FIG. 1, 1 is a microphone for inputting audio;
2 is an amplifier that amplifies only the audio band of the voice input from the microphone 1, and 3 is a feature extraction unit to which the output of the amplifier 2 is input.

The feature extraction unit 3 extracts feature parameters of a 16 ms section (hereinafter referred to as a frame) from the audio amplified by the amplifier 2 at intervals of 8 ms.
The feature parameters mentioned above are the one-syllable feature pattern (for example, output from a 16-channel band filter, etc.) used in the similarity calculation for final speech recognition performed by the matching unit 7, and the phoneme classification unit. 4 are parameters (for example, power, first-order autocorrelation coefficient, etc.) used for phoneme classification. The phoneme classification unit 4 uses the feature parameters to label one frame of the audio to indicate the nature of the audio of that frame. The boundary detection unit 5 extracts stable points on the spectrum by a method described in detail later, and detects syllable boundaries based on the stable points.

The syllable interval extraction unit 6 uses the time series of labels obtained by the phoneme classification unit 4 and the syllable boundary information obtained by the boundary detection unit 5 to extract syllable intervals from the input speech. Extract. Further, the matching section 7 performs the feature extraction section 3 in one syllable section extracted by the syllable section extraction section 6.
Speech recognition is performed by performing Euclidean distance calculation, which is an example of similarity calculation between the feature pattern extracted in step 1 and the feature standard pattern previously stored in the patent standard pattern memory 8. The CPU 9 includes the feature extraction section 3, the boundary detection section 5, and the syllable section extraction section 6.
In addition to controlling the matching section 7, it also controls an interface 10 that outputs the recognition results from the matching section 7 to an external device (not shown).

The speech recognition device having the above configuration operates as follows.

When an input person utters a voice into the microphone 1, the voice enters the microphone 1 and is amplified only in the voice band by the amplifier 2, and then sent to the feature extraction section 3.
The feature extraction unit 3 divides the frame into 16 ms frames at intervals of 8 ms, and extracts the feature parameters of the frames. The feature parameters include a one-syllable feature pattern (for example, output from a 16-channel band filter, etc.) used in similarity calculation for final speech recognition performed by the matching unit 7, and a phoneme classification unit 4. and parameters (eg, power, first-order autocorrelation coefficient, etc.) used for phoneme classification. In the phoneme classification section 4, a label representing the nature of the speech of the frame is performed using the feature parameters obtained by the feature extraction section 3. Here, there are four types of labels used in this embodiment: vowel (symbol 'V'), fricative (symbol 'F'), buzz bar (symbol 'B'), and silent (symbol '.').

In addition, the boundary detection unit 5 extracts a stable point from the obtained change in the spectrum of one frame, and further extracts the spectral pattern of the frame at the stable point and the spectral pattern of the audio frame input after the stable point. The syllable boundary of the syllable to be extracted is detected by calculating the Euclidean distance representing the degree of similarity between the syllable and the syllable. FIG. 2 shows a flowchart from extracting the stable point to detecting the syllable boundary. The right side of the figure is the flow for extracting the stable point, and the left side is the flow for detecting the syllable boundary.
The means for extracting the stable point and detecting the syllable boundary will be described in detail below with reference to FIG.

Here, each variable is defined as i, j: temporary variable, N: constant representing the order of the pattern, t: frame number, ta: frame number of the stable point, PAT(i): i-th order of the characteristic pattern at the stable point. Feature amount D(t): Spectral change distance at frame t, SP(t)(i): i-th feature amount of the input pattern at frame t, L: Constant representing the length of the window for calculating the spectral change. 2L+1 becomes the window length, M: A constant representing the length of the window for finding the stable point.
2M+1 is the window length, DIS: distance between the feature pattern of the stable point and the feature pattern of the input frame, ANTFLG: boundary detection flag based on the above distance from the stable point.

Now, one frame t on the spectrum
Input pattern from (let this be the current frame)
When SP(t)(i) is input, in step S ₁ , it is determined whether there is a stable point pattern (that is, whether a stable point has been extracted in the past and the spectral pattern of the above stable point has been incorporated). do. Here, we take advantage of the fact that if the spectral pattern of a stable point is captured, the data of the spectral pattern of a stable point will never become 0 in all orders, and we will calculate all PAT( When PAT(i)=0 is satisfied for i), the stable point pattern that has already been extracted is ignored and the process proceeds to step _S2 to begin the process of finding a stable point; otherwise, the stable point pattern that has already been extracted is assumed to be present. Proceed to step _S5 .

Step S ₂ checks the stability of the current frame. That is, if the spectrum change D(t) in the current frame t is D(t) = _N 〓 ⁱ⁼¹ (SP(t-L)(i)-SP(t+L)(i)) ² , then D(t) = min j ...(1) However, when there exists D(t) that satisfies j=-M, -M+1,,,0,,,M, the above current frame t
is judged to be stable, and the process proceeds to step _S3 . If there is no D(t) that satisfies the above equation (1), the current frame t is assumed to be unstable, and the process returns to step _S1 to execute the processing of the next frame.

In step _S3 , in order to avoid adopting the point where the spectral change is extremely large as the stable point,
The spectrum change D(t) in the stable frame t obtained in step _S2 is compared with the set value THDIS2. If the result is smaller than THDIT2, step
The process advances to step _S4 , and if the current frame t is above, it is determined that the current frame t cannot be adopted as a stable point, and the process returns to step _S1 .

In step _S4 , the feature pattern of the spectrum in the frame ta selected as the stable point is set as the stable point pattern PAT(i) to complete extraction of the stable point, and the process returns to step _S1 .

PAT(i)=SP(ta)(i) i=1,,N In step _S5 , the distance (DIS) between the stable point pattern of the stable points extracted above and the characteristic pattern of the spectrum in the current frame t is calculated as follows. Calculate using the formula DLS= _N 〓 ⁱ⁼¹ (PAT(i)−SP(t)(i)) ² and proceed to step _S6 .

In step S ₆ , calculate the distance DIS obtained in step S ₅ above.
is larger than the set value THDIS1, that is, whether the similarity is small or large. If it is less than the set value, the stable point pattern and the feature pattern in the current frame are similar, so the current frame is a syllable. It cannot be adopted as a boundary point, and the process returns to step _S1 . On the other hand, if it is larger than the set value, the current frame is determined to be a syllable boundary point and the process proceeds to step _S7 .

In step _S7 , when it is determined in step _S6 that IDS>THDIS1 and a syllable boundary is detected, the syllable boundary detection flag ANTFLG is set to ANTFLG=1 and the process proceeds to step _S8 .

Since the syllable boundary detection of the syllable to be extracted in step _S8 has been completed, the stable point pattern used for boundary detection is
Clear PAT(i) PAT(i) = 0 where i = 1, , N and return to step S ₁ to extract the stable point of the next syllable and detect the syllable boundary.

As described above, when a stable point of one voice is extracted and a syllable boundary of a syllable to be extracted is detected based on this stable point, the syllable segment extraction unit 6 of FIG. Based on the time series of syllable labels obtained in section 4 and the syllable boundary information obtained by the boundary detection section 5, the syllables are extracted by the syllable interval extraction section 6 according to the syllable extraction flowchart shown in FIG. Extracted.

Here, each variable is defined as SEG: label output by the phoneme classification section, FRAME: number of extracted phoneme frames, CUTFLG: extraction completion flag, ANTFLG: phoneme boundary detection flag, (detected by the syllable boundary detection section) FRMCNT: Frame counter, VCNT: Frame counter with vowel label 'V', THCUT: Constant (10) 'V': Vowel phonological label, 'F': Frictional phonological label, .

In step _S11 , it is determined whether or not CUTFLG (syllable extraction completion flag) has been set. If it has been set, the process proceeds to step _S12 , where the CUTFLG is cleared and the process proceeds to step _S13 . If you have cleared it, proceed directly to step _S13 .

In step _S13 , it is determined whether the SEG (phonetic label) of the current frame is 'V' or not. If 'V', the process proceeds to step _S14 , and if not 'V', the process proceeds to step _S17 .

FRMCNT (frame counter) at step S ₁₄
Add +1 to VCNT (vowel phonological label′)
Add +1 to the frame number of V' and proceed to step _S15 .

In step _S15 , whether the ANTFLG (syllable boundary detection flag) is set or not (this ANTFLG
is set when the detection of one syllable boundary is completed in step _S7 of the flowchart of stable point extraction and syllable boundary point detection in FIG. ).
As a result, when set to 1, the step
Proceeding to _{step S16} , one syllable is extracted, and if it is not set, it is determined that one syllable boundary detection has not yet been completed, and the process returns to step _S11 to execute the processing of the next frame. At step ₁₆ , follow the steps above.
When it is determined in _S15 that the above ANTFLG is set to 1, a boundary of one syllable has been detected, so the frame up to the current frame is regarded as one syllable, and the number of syllable frames up to the current frame is counted. Transfer the above FRMCNT to FRAME (frame number of extracted syllables), clear the above FRMCNT and above VCNT, set the 1 syllable extraction completion flag CUTFLG to 1, return to step _S11 , and proceed to the next step. Execute syllable extraction processing.

At step _S17 , the phonetic label of the current frame is
When it is not V′, the above VCNT and THCUT (constant =
In this example, 10) will be compared. As a result, if the number of vowel phonological labels is greater than THCUT,
It is determined that the frames before the current frame are meaningful syllables and that the current frame is a syllable boundary, and the process proceeds to step _S18 to extract one syllable.
If it is less than THCUT, proceed to step _S19 .

In step _S18 , the FRMCNT, which counts the number of frames of syllables up to the current frame, is counted as one syllable, and the FRMCNT and VCNT are cleared, and one syllable extraction is completed. The flag CUTFLG is set to 1 and the process returns to step _S11 .

In step _S19 , it is determined whether the above-mentioned SEG of the current frame is 'F' or not. If 'F', the process proceeds to step _S21 , and if not 'F', the process proceeds to step _S20 .

In step _S20 , if the SEG of the current frame is neither 'V' nor 'F', it is assumed that the syllables up to the current frame are not meaningful syllables, the above FRMCNT and VCNT are cleared, and the process returns to step _S11 . Execute the syllable extraction process.

In step _S21 , when the phoneme label is 'F', it is assumed that syllables are still continuing, so +1 is added to the execution FRMCNT, and the process returns to step _S11 to execute the processing of the next frame.

Figure 3 Steps of syllable extraction flowchart
When the one-syllable extraction completion flag CUTFLG is set to 1 in _S16 and step _S18 , the matching section 7 performs the following in accordance with a command from the CPU 9 shown in FIG.
Calculate the degree of similarity between the feature pattern of one syllable section extracted by the syllable section axis extraction unit 6 of the input speech and the feature standard pattern stored in advance in the feature standard pattern memory 8. The input and extracted syllable is recognized as the same syllable as the standard syllable with the highest degree of similarity, and the recognition result is output to an external device via the interface 10.

Figure 4 shows the stable points extracted in this example.
An example of syllable boundary points is shown, and from the top, a time series of phoneme classification labels, a vowel series (reference) obtained by a method different from this example, and spectral changes are described. Further, C represents the syllable boundary point determined from the conventional spectrum change, and A and B represent the syllable boundary points determined in this embodiment. Note that, from FIG. 4, since the phoneme classification label is 'V' for all vowels, FIG. 4 is an example in which a syllable boundary is detected in step _S15 in the syllable extraction flowchart of FIG. 3.
That is, a stable point P ₁ is set on the above spectrum curve by the above method, and based on this stable point P ₁ , the distance DIS between the characteristic pattern of each frame and the above stable point pattern is determined by the above method as shown in the figure. thick curve
It is calculated as P ₁ Q ₁ , and at point Q ₁ , DIS＞
The result is THDIS1, and syllable boundary point A is detected. Similarly, when the next stable point P ₂ is set, point Q ₂ is found based on P ₂ , the next syllable boundary point B is detected, and the three syllables ``e'', ``i'', and ``o'' are ' is separated and extracted. In the conventional method of detecting syllable boundaries from spectral changes, the extreme points of spectral changes are
Since the syllable boundary point C is detected from P ₃ , the syllables "ei" and "o" are distinguished and extracted, but the syllables "e" and "i" have a gentle spectrum change between the two syllables. Therefore, syllable boundary points are not detected, and therefore, it is not possible to distinguish and extract different syllables.

Therefore, in this embodiment, even when the spectral change is small and it is impossible to extract the syllable boundary using the conventional spectral change, the syllable boundary can be detected accurately.

<Effects of the Invention> As is clear from the above, the syllable recognition device of the present invention includes a stable point extraction unit that extracts a stable point of the spectral information from a change in the spectral information in the input speech following the extracted speech section; Next, the similarity is calculated between the spectral information at the stable point extracted above and the spectral information after the stable point, and the calculated similarity value is compared with a predetermined value. Since a syllable boundary detection unit is provided to detect the syllable boundary of the syllable interval, the syllable boundary can be detected accurately and easily even when the change in the spectral information is gentle or sudden. be able to.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the speech recognition device of the present invention, FIG. 2 is a flowchart of stable point extraction and syllable boundary detection, FIG. 3 is a flowchart of syllable extraction, and FIG. 4 is a stable point extraction and syllable boundary detection flowchart. It is a figure which shows an example of a point and a syllable boundary point. 3...Feature extraction unit, 4...Phonological classification unit, 5...
Boundary detection unit, 6...Syllable section extraction unit, 7...Matching unit, 8...Characteristic standard pattern memory, 9...
…CPU.

Claims (1)

[Scope of Claims] 1. A syllable section is extracted from input speech, and the degree of similarity between the feature pattern of the extracted syllable and a feature standard pattern stored in advance in memory is calculated, and the input speech is extracted. A speech recognition device that recognizes syllables in units of syllables includes: a stable point extraction unit that extracts stable points of the speech spectrum information from changes in the speech spectrum information in the input speech following the extracted syllable section; The similarity between the speech spectrum information at and the speech spectrum information after the stable point is sequentially calculated, and the calculated similarity value is compared with a predetermined value to determine the syllable boundary of the next syllable interval to be extracted. A speech recognition device comprising: a syllable boundary detection unit for detecting a syllable boundary.