JPS6370899A - Voice recognition equipment - Google Patents

Abstract

(57) [Abstract] 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 speech recognition device that recognizes input speech such as Japanese in syllable units and outputs it to an external device.

<Prior art> Conventional speech recognition devices use changes in the speech spectrum (hereinafter simply referred to as spectrum) in a certain interval in order to extract the syllable interval of the syllable, which is the recognition unit, from input speech. The boundary between the syllables is detected using the syllable.

<Problems to be Solved by the Invention> However, in the above-mentioned conventional speech recognition device, changes in the spectrum of the input speech include a mixture of rapidly changing speech and gently changing speech. It is difficult to accurately detect syllable boundaries by following the syllable boundaries, and there is a problem that speech misidentification often occurs.

SUMMARY OF THE INVENTION 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 how abrupt the spectrum changes.

<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 advance in memory. The speech recognition device' calculates the degree of similarity with the stored characteristic standard pattern and recognizes the input speech in syllable units. A stable point extraction unit that extracts a stable point calculates the degree of similarity between the spectral information at the stable point and the spectral information after the stable point, and compares the degree of similarity with a predetermined value to determine the point to be extracted. The present invention is characterized in that it includes a syllable boundary detection section that detects syllable boundaries in syllable intervals.

<Operation> When audio is input, the stable point extraction section extracts a stable point on the spectrum using changes in spectral information in a certain section of the input audio. Furthermore, the syllable boundary detection unit calculates the similarity between the spectral information at the stable point and the spectral information after the stable point, and compares the obtained similarity with a predetermined value. A syllable boundary of a syllable interval is detected, and a syllable interval is extracted according to the detected syllable boundary. Therefore, meaningful syllable sections can be extracted accurately regardless of the degree of change in the spectrum.

<Example> Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments.

In FIG. 1, reference numeral I designates a microphone that inputs audio, 2 an amplifier that amplifies only the audio band of the audio input from the microphone 1, and 3 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 m5 section (hereinafter referred to as a frame) of the audio amplified by the amplifier 2 at intervals of 8 ms. The feature parameters mentioned above are the l-syllable feature pattern (for example, output from a 16-channel bandpass filter, etc.) used in the similarity calculation for final speech recognition performed by the matching unit 7, and the feature pattern by the phoneme classification unit 4. These 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 using a method described in detail later, and detects syllable boundaries based on the stable points.

The syllable interval extraction unit 6 extracts syllable intervals from the input speech using the time series of labels obtained by the phoneme classification unit 4 and the syllable boundary information obtained by the boundary detection unit 5. Furthermore, the matching section 7 matches the feature pattern extracted by the feature extraction section 3 in one syllable section extracted by the syllable section extraction section 6 with the feature standard pattern stored in advance in the feature standard pattern memory 8. Speech recognition is performed by performing Euclidean distance calculation, which is an example of similarity calculation. The CPU 9 controls the special first place extraction section 3, the boundary detection section 5, the syllable section extraction section 6, and the matching section 7, and also provides an interface for outputting the recognition results obtained by the matching section 7 to an external device (not shown). It controls 10.

Mizonari's speech recognition device operates as follows.

When an input person utters a voice into the microphone 1, the voice enters the microphone 1, is amplified only in the voice band by the amplifier 2, and is sent to the feature extractor 3. Each interval is divided into frames of t6ns, and the characteristic parameters of the frames are extracted.

The above feature parameters are determined by the matching unit 7. The feature pattern of l syllables (for example, output from a 16-channel bandpass filter, etc.) used for similarity calculation for final speech recognition, and the phoneme classification unit 1. These are parameters used for phoneme classification (eg, power, first-order self-use relation coefficient, etc.). In the phoneme classification section 4, the above-mentioned characteristic extraction section 3 performs labeling that represents the nature of the speech of the frame using the feature parameters. Here, the labels used in the untrue case are vowel (symbol 'V'), fricative (symbol 'F'), buzzbar (symbol '
There are four types: B'), silence (symbols ',').

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 the extraction of the stable points to the detection of syllable boundaries, in which the right side of the figure is the flow of stable point extraction and the left side is the flow of syllable boundary detection. 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 i, j is a temporary variable, N is a constant representing the order of the pattern, t
: frame number, ta: frame number of stable point, PA'r(i): first-order feature amount of feature pattern of stable point D (t) spectrum change distance at frame t, S P (tXi): frame The first-order feature of the input pattern at t, L is a constant that represents the length of the window for calculating the varnish vector change, and 2L+1 is the window length. ~I: The constant that represents the length of the window for finding a 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, ANTPLG: boundary detection flag based on the above distance from the stable point.

Now, the input cover turn 5P(t)(i) from a certain frame t on the spectrum (this is the current frame)
When is input, in step SI, it is determined whether there is a stable point pattern (that is, whether a stable point has been extracted in the past and the spectrum pattern of the stable point has been imported). Here, i=1. , N, when P,AT(i)=0 is satisfied, there are no stable point patterns that have already been extracted, and the process proceeds to step S2, where the process begins the operation of finding stable points. In other cases, it is determined that there is a stable point that has already been extracted, and the process proceeds to step S5.

In step S, the stability of the current frame is checked.

That is, the spectral change D (
j) as D (t)−Σ(S P (t−LXi) −S P
(t+LXi))"i=1, D(t)-min D(t-I-j)・1 @・(
1) However, J” M, M+1 ,,O,,,
When D (t) that satisfies M exists, the current frame t is judged to be stable and the process proceeds to step S3, and the above (+)
If there is no D(t) that satisfies the equation, the current frame t
is not stable, and returns to step S1 to process the next frame.

In step S, in order to avoid adopting a point where the spectrum change is extremely large as a stable point, step S2
The spectral change D (t) in the stable frame t determined in is compared with the set value THDIS2. As a result, if it is smaller than THDIS2, the process proceeds to step S4, and if it is greater than THDIS, it is determined that the current frame t cannot be adopted as a stable point, and the process returns to step S.

In step S4, the frame ta adopted as the stable point
The characteristic pattern of the spectrum in is set to the stable point pattern FAT(i), the stable point extraction is completed, and the process returns to step S.

FAT(i)-SP(taXi) 1=11. In step S, the distance #(DIS) between the stable point pattern of the extracted stable points and the characteristic pattern of the spectrum in the current frame t is calculated using the following formula, and step S,
Proceed to.

In step S6, it is determined whether the distance DrS obtained in step S5 is larger than the set value THD I S 1, that is, whether the similarity is small or large. If it is less than the set value, the stable point pattern and the current frame are determined. Since the current frame is similar to the feature pattern, it is determined that the current frame cannot be adopted as a syllable boundary point, and the process returns to step SI. 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 advances to step S7.

At step S, at step S, DIS>THDIS+
When it is determined that the syllable boundary is detected, the syllable boundary detection flag ANTFLG is set ANTFLG=
1 and 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 F A T used for boundary detection has been completed.
Clear (i) FAT(i)-〇 where i=I, . . . N and return to step SI to extract the stable point of the next syllable and detect the syllable boundary.

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

Here, each variable is SEC: label output by the phoneme classification section, FRAME
: Number of extracted syllable frames, CUTFLG: Extraction completion flag, ANTFLG': Syllable boundary detection flag, (detected by the syllable boundary detection unit) PRMCNT: Frame counter, VCNT
: Counter of frames with vowel label /V/, THCUT : Constant (10) '■' : Vowel character phoneme label, 'F' : Fragile phoneme label.

In step S11, it is determined whether or not CUTFLG (syllable extraction completion flag) is set. If it is set, the process proceeds to step SI2, where the cUTFLG is cleared and the process proceeds to step S13. - If it has been cleared, proceed directly to step SI3.

In step Sij, 5EG (phonological label) of the current frame
Determine whether or not is /V/, and if /V/, step S
The process proceeds to step S14, and if it is not /V/, the process proceeds to step S+7.

In step S14, FRMCNT (frame counter)
+1 is added to VCNT (the number of frames of the vowel phonetic label '■'), and the process proceeds to step S+5.

In step S +5, it is determined whether the ANTFLG (syllable boundary detection flag) is set (this ANTFLG indicates that one syllable boundary detection has been completed in step S7 of the stable point extraction and syllable boundary point detection flowchart in Figure 2). ). As a result, if it is set to l, the process advances to step S141 to extract l syllables, and if it is not set, it is determined that the boundary detection of l syllables has not yet been completed and the process returns to step S11. Executes processing for the next frame. Step SI
ll, when it is determined in step SI5 that the above ANTF'LG is set to the above mentioned ANTF'LG power moon, the l syllable boundary has been detected, so the syllable up to the current frame is regarded as one syllable, and the syllable up to the current frame is Transfer the above PRMCNT counting the number of frames to FRAME (frame number of extracted syllables), clear the above FR) ν1CN T and the above VCNT, and set the l syllable extraction completion flag C.
UTFLC; is set to l and the process returns to step S I+ to execute the next syllable extraction process.

In step SI7, when the phonetic label of the current frame is not 'V', the above VCNT and THCUT (constant = 10 in this embodiment) are compared. As a result, if the number of vowel phonetic labels is greater than THCUT, it is determined that the frames before the current frame are meaningful syllables and the current frame is a syllable boundary, and the process proceeds to step S1B. 1
Perform syllable extraction, and if it is less than THCUT, step SI
Proceed to step 8.

In step S18, the FRMCNT counting the number of frames of syllables up to the current frame is regarded as one syllable, and the FRMCNT is transferred to the FRAME, and the FRMCNT and VCNT are transferred to the FRAME. The flag CUTFLG indicating completion of l-syllable extraction is set to 1, and the process returns to step S11.

In step S19, the above SEG of the current frame is 'F'
If it is 'F', proceed to step S21; if not 'F', proceed to step S. Proceed to.

Step S. If the SEG of the current frame is neither /V/ nor /P/, the syllables up to the current frame are not meaningful syllables, and the above F RM CNT and V
The CNT is cleared and the process returns to step S II to execute the next 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 execution FRMCNT, and the process returns to step SII to execute the processing of the next frame.

At step S+e and step Sla of the syllable extraction flowchart in FIG. 3, the l syllable extraction completion flag CUT is set.
When FLG is set to 1, the matching unit 7, in response to a command from the CPU 9 shown in FIG. The degree of similarity with the characteristic standard pattern stored in advance in the standard pattern memory 8 is calculated, and the input and extracted syllable is recognized as the same syllable as the drifting syllable with the highest degree of similarity, and the recognition result is 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. As shown in FIG. 4, all the phoneme classification labels are vowel /V/, so FIG. 4 is an example in which a syllable boundary is detected at step S+5 in the syllable extraction flowchart of FIG. 3.

That is, a stable point P is set on the spectrum curve by the above method, and based on this stable point P, the distance DIS between the feature pattern of each frame and the stable point pattern is determined by the above method as shown in the figure. is determined as a large curve P1Q+, and at point Q1, DIs>THDISI and syllable boundary point A is detected. Similarly, the next stable point P,
Once set, point Q2 is found based on P, the next syllable boundary point B is detected, and the three syllables "e", "i", "
"O" is separated and extracted. In the conventional method of detecting syllable boundaries from changes in the spectrum, the syllable boundary point C is detected from the extreme point P3 of the spectrum change, so the syllables "ei" and "o" are extracted separately, but the syllables "Uh"
Since the spectral change between the two syllables of "I" is gentle, the syllable boundary point is not detected, and therefore it is not possible to distinguish and extract the syllable as a different syllable.

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 section that extracts a stable point of the spectrum from changes in spectral information in a certain section of input speech; The syllable boundary detection gokuto detects the syllable boundary of the syllable interval to be extracted by calculating the similarity between the spectral information at and the spectral information after the stable point and comparing the similarity with a predetermined value. Since the setup is robust, syllable boundaries can be detected accurately and easily even when the change in the spectral information is gradual or sudden.

[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 shows the stable points extracted in the embodiment. It is a figure which shows an example of a syllable boundary point.

Claims (1)

[Claims] (1) Extract syllable intervals from the input speech, calculate the similarity between the extracted syllable feature pattern and the feature standard pattern stored in memory in advance, and extract the input speech syllable by syllable. a stable point extraction unit that extracts a stable point of the voice spectrum from changes in voice spectrum information in a certain section of the input voice; A syllable boundary detection unit that calculates the degree of similarity with the speech spectrum information after the point and compares the degree of similarity with a predetermined value to detect the syllable boundary of the syllable interval to be extracted. speech recognition device.