JPS6289999A - Voice analyzer - 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 [Field of Industrial Application] The present invention relates to a speech analysis device that analyzes an input speech signal based on its spectral envelope.

``Conventional technologies such as speech recognition devices, pronunciation training devices for hearing-impaired people, communication systems and speech synthesis devices that use speech analysis and synthesis require the analysis and analysis of input speech signals in order to achieve the desired processing. Feature extraction is required.The analysis of the input audio signal is typically done based on its frequency spectrum.

This is because human hearing is more sensitive to the spectrum of an audio signal than to the time-varying waveform itself, and recognizes signals with the same spectral shape as the same phoneme. It depends on the facts.

The voice signal has a structure as a periodic signal driven by the vocal cord t5 movement in the voiced part. As a result, the frequency spectrum of the voiced part has a line spectral structure. On the other hand, in the silent portion, the voice signal is not accompanied by vocal fold vibration, but rather has its source as noise caused by airflow passing through the vocal tract. As a result, the frequency spectrum of this silent portion does not have a periodic structure like a line spectrum. Correspondingly, in conventional speech analysis, there are methods that assume a periodic pulse generation source as the sound source of the input audio signal, and methods that assume a noise source. The former is an autoregressive model (△R model)
The latter is well known for its speech analysis using cepstral analysis. According to these audio analyzes, a so-called spectral envelope obtained by removing fine structures from the spectrum of the input audio signal is obtained.

When analyzing an input audio signal using a method such as the AR model analysis method or cepstral analysis method described above to obtain the spectral envelope, these methods assume the temporal stationarity of the system, so inherently , cannot be applied when the phoneme changes over time. Therefore, in these analysis methods, the signal is cut out at short intervals where the system does not appear to change significantly, and a window function such as a Hamming window or a Hanning window is applied to this signal to avoid the influence of the end points. By applying this, a temporally quasi-stationary signal is created. The spectral envelope obtained by analyzing this signal by 1 is taken as the spectral envelope at the time of signal extraction.

In addition to the analysis method for obtaining a time series of the spectral envelope described above, a method for analyzing non-stationary audio signals that assumes temporal changes in the system as a premise of the model has also been proposed. However, even with this analysis method, a time series of spectral envelopes is still obtained as a result.

Furthermore, as a powerful analysis method other than the above-mentioned analysis method, there is a frequency analysis method using a filter bank. In this analysis method, an input audio signal is passed through a number of band-pass filters each having a different center frequency, and the outputs of these filters are used as the spectral intensity. A feature of this analysis method is, for example, the ease of processing between actual and old data using hardware.

In this way, in speech analysis, the spectral envelope of an input speech signal is often determined, but the final method of analyzing the speech signal from the obtained spectral envelope is to analyze its local peaks (hereinafter referred to as local peaks). Formant analysis is known in which analysis is performed by extracting formant frequencies, widths, etc. from This analysis method is based on the fact that each vowel has its own 7-orman l-frequency and formant width, and that each consonant is characterized by the way the formant frequency changes toward the following vowel. It is based on For example, the five Japanese vowels (iweo) are colored by two formant frequencies F+ and F2 from the lowest frequency side, and in the voices of people of the same gender and age, Fl and F2 have almost the same values. Therefore, Forman (−
Frequency Fl.

By detecting F2, it is possible to identify the vowel.

Also, local peak analysis is known, which extracts the local peak of the spectral envelope and focuses on its frequency and temporal transition, without worrying about formants. This analysis method is based on the fact that phonological features are thought to appear in temporal changes in local peaks in vowel changes and consonant parts.

In addition, a method has been proposed in which the spectral envelope curve itself is used as the feature quantity of the audio signal and used for identification, classification, or display in subsequent processing.

As mentioned above, in the analysis of audio signals, it is important to extract the spectral envelope, and in addition to the spectral envelope itself, it is important to extract the formant frequency and width obtained from it, as well as the local peak. The frequencies and their transitions are also used to characterize speech.

[Problems to be Solved by the Invention] By the way, when a human utters a voice, the phonology is thought to be formed by the resonance and anti-resonance characteristics of the vocal tract. For example, resonant frequencies appear as formants on the spectral envelope. Therefore, in humans with almost the same vocal tract structure, a wide range of spectra can be obtained for the same pitch.

However, it is known that when the vocal tract lengths are significantly different, such as between men and women, children and adults, the resonance or antiresonance frequencies shift, and the shapes of the spectral envelopes for the same phoneme do not match. There is. Therefore, in this case, the local peak and formant frequencies will also shift. This is useful for analyzes that aim to extract the same results for the same phoneme regardless of the speaker, such as speech recognition for unspecified speakers or visual display of speech for the hearing impaired. is extremely inconvenient.

Conventionally known methods for solving this problem include preparing a large number of standard patterns and calculating the ratio of formant frequencies.

In the former method, the spectral envelopes of many different people, such as men, women, adults, and children, are registered as 1111% patterns, and unknown input patterns are classified into the most similar one among the many standard patterns. By doing so, it attempts to recognize an unspecified number of input voices. However, this method has the disadvantage that a large number of standard patterns must be prepared in order to handle any human input voice, and it takes a long time to compare these patterns. be. Furthermore, since this method does not extract the result of normalizing the vocal tract length, it cannot be used for the purpose of displaying phoneme features that are not dependent on vocal tract length.

The latter method of taking the ratio of formant frequencies is well known as a method for extracting features of vowels that are independent of vocal tract length. For further explanation of this method,
First, among the local peaks in the spectrum envelope, the first peak is considered to be relatively stable for vowels. Second. Third
Extract the frequencies F1, F2, Fl of the fol7 tones and calculate the ratio between them, e.g., Fl/F3, F2/F
3 is obtained and used as a feature quantity. If the vocal tract length increases by a times, the formant frequency increases by i/a times, that is, Fl
/a, F2 /a, Fl /a Nina Luka,
The basis of this method is that the ratio of solera remains unchanged.

This method uses vowels 'M1. 1st. If the third formant is stably extracted, good results are obtained, but if it is not stably extracted, there is a problem in that the reliability of the analysis results is significantly reduced. Furthermore, this method has the disadvantage that it cannot be applied to consonant parts. In other words, in the consonant part, the formant, which is a syngeneic characteristic of the vocal tract, is not defined, and in fact, the first . Second
．． Since the local peak corresponding to the third formant is not always observed, it is not possible to extract Fl, F2, and Fl and calculate their ratio. In addition, formants are not necessarily stable not only in consonant parts but also in vowel parts, such as the rising part and the ending part, so that incorrect formant frequencies may be extracted. In such a case, the formant frequency ratio changes discontinuously and takes a completely incorrect value. Therefore, this method can only be used for stable vowel parts in a speech signal, and the beginning and end of vowel parts and consonant parts must be analyzed using another method. . However, in that case, the extraction parameters are different between the ffi sound stable part and the other parts, so it is not possible to record a continuous change from the consonant part to the vowel part. In other words, the method of taking the ratio of Forman 1 to frequency is essentially applicable only to stationary vowels.

As described above, a method for extracting phoneme-specific features ff1l from the spectral envelope of general speech of an unspecified number of voices with different vocal tract lengths has not yet been found.

The purpose of this invention is to
! To provide a speech analysis device capable of obtaining a spectral pattern specific to phonemes without being influenced by differences in speech.

[Means and operations for solving the problem] In order to achieve the above object, the present invention logarithms and normalizes the spectral envelope extracted from the input audio signal, and the normalized logarithmic spectral envelope obtained thereby. is integrated on the Mel scale, and the envelope information of the spectrum envelope is projected from the frequency onto the peak output.

[Embodiment 1] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment. Before explaining FIG. 1, an outline of the embodiment will first be explained with reference to FIGS. 2 to 7.

First, a comparison of the spectral envelopes of vowels depending on vocal tract length is shown in Figure 2. Figure 2 shows two different vocal tract lengths (
J, t >, (J12) is a logarithmic representation of the spectral envelope Pi) of the same sound ■. According to this Figure 2, in the frequency range from several hundred Hz to about 5 kHz, the spectral envelope P of a long vocal tract length of 1M+)
t(f) (in Fig. 2, logPt(J)) is the spectral envelope P2 (J) (second
In the figure, the spectral envelope Pi) having approximately the same shape as l0TJ F2 1 )) is multiplied by a constant in the frequency axis (f> direction) with the origin fixed.

On the contrary, from OHz to several hundred Hz and 5 to 2 days
In a frequency range above a certain degree, differences related to individual sound quality become significant, and the similarity of the spectral envelopes Pi) becomes small. In other words, vocal tract length (Jlt >, (J!,
2 > ratio 1! t/, 1.2> is r, then the difference between the spectral envelope P1 (J> and the spectral envelope Pz (J>) is 110 when the magnitude is normalized and the logarithm is taken.
From 0Hz to 5kHz and in degrees Pii, the following relationship holds true: (1). Here, the spectral envelope P+ (
J), F2 (H) itself, but to remove the difference in the amplitude of the thick voice signal.

Now, at this time, if the first to third formants are extracted, their frequencies (Fl) and (F2) are obtained.

(F3) and (F1'), (F2!'), (F3'
) is shown in Figure 2. In this case, between these
Since there is a relationship of zF3'/F3tr (2), the ratio of formant frequencies (F) (for example, as shown in the following equation (3)) remains unchanged.

Fl　・'F2 Evening Fl', /F2'.

Fl, 'F3ri Fl', /F3' ... (
3) This will be explained according to the actual measurement examples shown in FIGS. 3 and 4. Figure 3 shows the distribution of Fl and F2 for men and women in their 20s to 30s for the five Japanese vowels. As is clear from Figure 3, the distribution of men and women is significantly different.
Male workers and female workers are distributed in the same area.

In contrast, Fig. 4 shows Fl /F3, F2 /F
This shows the distribution of 3. According to this Figure 4,
It can be seen that the difference between men and women has disappeared in the ratio of Forman 1 to frequency.

Therefore, not only the constant part of vowels, but also frequencies from several hundred Hz to about 5
In the range of about kHz2 [constant times the frequency axis, 1-r
Transformation of the spectral envelope 1flP()) by -t (R) P(J)→P' 1)=P(r i) -+4+
If we find a transformation (U) that projects the spectral envelope P(f>) onto an invariant functional space, then in that space, the spectral envelope P(J) belonging to the same phoneme, regardless of the vocal tract length (1), can be found. should have the same shape.

FIG. 5 conceptually shows this. This figure 5 shows that the difference in the spectral envelope P(f) of the sound fffi "a" or "i" due to the difference in vocal tract length (1) is caused by the transformation (U
) into a spectral envelope P'(f) of the same shape.

That is, for the phoneme A'', the spectral envelope P1 at the vocal tract length 11) and the spectral envelope P27 (f) at the vocal tract length (12) are transformed into a spectral envelope P' of the same shape by transformation (U). ?(+).Similarly, regarding the phoneme I, the spectral envelope P1 I<f>.

This figure shows how P2 i<f) is converted to the spectrum envelope P' i(ji).

In this embodiment, the above conversion (Ll) is realized as follows. In other words, the integral of the logarithmic spectral envelope whose size has been normalized over the Mel scale ~1(f) is L(chi)
Then, L(J)=f', log IP(k>ldM(
k) ...(5). Here, Mel scale M(f>
is known as a measure of the sense of pitch in human hearing, and it also coincides with the relationship to frequency of auditory cells distributed on the basal model of the octopus canal, which is the auditory organ. It has been known. In other words, it is estimated that humans perform frequency analysis of speech on the Mel scale M (1).

FIG. 6 is a graph showing the dependence of the Mel scale M(7>) on frequency, and its functional form is given by, but in the frequency range of about 500+-(Z to 5kHz), it can be approximated as follows.

Since Li) in equation (5) depends on the functional form of P(ji), it is written as Lp(J). If we apply the transformation of equation (4) to this, we get (where h=rk)...(8) However, in the frequency range we are currently considering (several 100 Hz to about 5 kHz), from equation (7), , d M (-) evening d
M (h) ...■... (10
), and Lp ' (7) Lp (+' f ) −
(11) holds true. Therefore, the spectral envelope P(ヂ)
Considering a function (P(J), Lp (chi)) in which p(+) and p(+) are expressed as parameters by frequency <i>, (P(J), Lp (()) = (P (r f ), Lp (r
) ) = (P' (J), Lp'(f>
) ...(12), (P (J), f> - (P <f>,
It can be seen that the transformation (tJ) of L ρ (f ) )...(13) is a projection onto the function space that is invariant with respect to the transformation (R) of equation (4).

FIG. 7 is a conceptual diagram of conversion (U). Note that FIG. 7 shows the spectral envelope P (not J>, but the logarithmic spectral envelope 1oqlPi) I as envelope information to be described later. In this case, the above equation (12) is expressed as the following equation (14).

(log IP (J)I, Lp (f)) - (log
IP' (J)I, LP'<i>>...(1
4) Also, the spectral envelope P<f> and the logarithmic spectral envelope 1
If we use the normalized logarithmic spectral envelope 100 I P < J) 1 instead of ooIP(J)I, then the equation (12
) is expressed as in the following equation (15).

...(15) The outline of this embodiment has been explained above, and next, returning to FIG. 1, the configuration for performing the actual addition process will be explained.

In FIG. 1, reference numeral 11 is a spectral envelope extraction unit that extracts the spectral envelope tgP(l) of the input audio signal (, A+N). Various methods can be used, such as an envelope extraction method, a spectral envelope extraction method in a speech analysis method using kebstral analysis, and a spectral envelope extraction method in a speech analysis method using frequency analysis using a filter bank.

12 is a logarithmization unit that logarithms the magnitude of the spectrum envelope P(l) extracted by the spectrum envelope extraction unit 11.

13 is a normalization unit that normalizes the magnitude of the logarithmic spectrum envelope 1oalP(f)I output from the logarithmization unit 12. Here, the logarithmic spectral envelope is 10! :1IP
i) As a method for normalizing the magnitude of I, for example, a method using automatic gain control, a method using logarithmic spectral envelope 1oalP
i) After differentiating 1 with the frequency (f) and dropping the constant term,
Examples include a method of integrating and adding a constant value.

14 is the normalized logarithmic spectral envelope 181og IP(f)I output from the normalization unit 13 using the mel scale M(
This is an integration section that integrates f) as a variable.

In other words, this integrating section 14 has a spectral envelope of 10QIP
(f) l is integrated according to the integral function shown in equation (5) above.

15 receives the logarithmic spectrum envelope 1oolP(J)I output from the logarithmization unit]2 and the integration result output from the integration unit 14, and calculates the logarithmic spectrum envelope logIP(J
) I onto the integral function L(f>(=Lp (f)) from the frequency <7> as shown in FIG. 7 above, and displays the projection result. In other words, , this projection unit 15 takes Lp (f) on the y-axis of the x-y orthogonal coordinates, takes the logarithmic spectrum envelope log I P l )l on the y-axis, and expresses it as a parameter by frequency <i>. This is to pattern the analysis results of the input audio signal (AlN).

In addition, in the processing of the projection unit 15, the values taken on the y-axis include the spectral envelope P(f) and the normalized logarithmic spectrum, as is clear from the explanation of equations (14) and (15) above. It is not limited to the envelope, but may be the normalized spectral envelope #'dP<f>. That is, in this invention, the envelope information provided for projection means at least the four patterns described above.

In addition, in the processing of the projection unit 15, it goes without saying that the envelope information may be plotted on the y-axis and Lp(f) may be plotted on the y-axis.

Here, an example of actual measurement using voice analysis in this embodiment will be explained.

Figures 8 and 9 show the female and male phonetic pie°', respectively.
It is the logarithmic spectrum envelope of .

First, the audio signal (AIN) input to the extraction unit 11 is
50 μse of audio collected with a condenser microphone
It is sampled at a period of c and digitized with 12 bits.

Note that Word's Uniibu memory is used in 3 to collect this audio.

The extraction unit 11 obtains the spectral envelope Pi) from this input audio signal <A, N) by cepstral analysis. The cepstral analysis is set to obtain the spectral envelope P(+) by differentiating the 1024-point frame of the vowel stable part, applying a Hamming window to this, and Fourier transforming it using the FFT algorithm. .

The logarithmization unit 12 takes the logarithm of the absolute value of the spectrum envelope P(di) obtained as described above, performs an inverse Fourier transform on it to obtain a cepstrum, and then converts it to a cutoff of 1.7 on the que frency. A logarithmic spectral envelope togI P l()l is obtained by applying a rectangular window of ~2.5 m sec and Fourier transforming the window.

In addition, this logarithmic subvek i・le envelope 1ogI P (J
)l, the cutoff on the above-mentioned quefrency is selected in accordance with the pitch frequency. Furthermore, in order to normalize the size of the logarithmic spectrum envelope 10qlPi)l, this spectrum envelope 1ooIPi)I is obtained after adjusting the O-th winding component of the cepstrum to a constant value.

The logarithmic spectral envelope 1oalP (
[)1 is obtained as described above, but if we compare the logarithmic spectral envelopes in Figures 8 and 9, we find that the spectral shapes of both are similar in the frequency range below about 5kl-12. However, in the frequency range beyond this,
The spectral shape of women is shifted higher.

For this logarithmic spectral envelope 1oglPi)I, Lp(J) expressed by equation (5) is obtained, and this is taken on the X axis, and the logarithmic spectral envelope 1ogIP() is taken on the y axis. 10 and 11. Looking at these, it can be seen that although the peak heights and fine nuances are slightly different, the deviation in the frequency direction between FIG. 8 and FIG. 9 has been eliminated.

FIG. 12 shows the logarithmic spectrum envelope 101JIP(f)I as described above while temporally shifting the frame position with respect to the rising part of the phonetic pattern uttered by a woman.
By finding this spectral envelope too I
It shows the time-series change of P (chi)1. 1st
Figure 3 shows the logarithmic spectrum envelope 1og 1Pi of Figure 12)
By applying the transformation (U) of this example to 1,
This shows the time-series changes in the voice analysis results.

According to this actual measurement example, it can be seen that the conversion (U) of this embodiment acts stably on consonant parts and vowel rising parts, and spectral changes are displayed smoothly.

Figures 14 and 15 show the time-series changes in the phoneme a'' for men and women, respectively, obtained in the same manner as in Figure 13. Looking at the vowel part in the figure, it can be seen that the large difference in vocal tract length between men and women has been eliminated.

As detailed above, in this embodiment, the envelope information (P(J)
, P(+), log IP(J)l or +oalP
([)I) from above frequency (f) by transformation (U) defined by equations (5) and 9. ＞By projecting it upward, the result of speech analysis is obtained.

According to such a configuration, it is possible to eliminate differences in analysis results due to differences in vocal tract length (Jl) of speakers, and always
It is possible to obtain phoneme-specific analysis results. In this case, this embodiment can be applied to the spectral envelope of any part of the input audio signal (AIN), regardless of whether it is voiced or unvoiced or whether it is a vowel or a consonant. Moreover, since the analysis result does not depend on the extraction accuracy of the formant frequency (F) or its stability, it can be applied to all parts of the input audio signal (AlN). In particular, according to this embodiment, the spectrum in the process of transitioning from the consonant part to the vowel part is not affected by differences in individual vocal tract lengths (1), which was completely impossible in the past. The problem of finding changes in envelope can be solved.

In addition, in this embodiment, the spectral envelope P(J) and the logarithmic spectral envelope 1 are used as the integrand of equation (5).
Since the normalized logarithmic spectral envelope loc+1Pi)I is used instead of o(IIP(f)l), the influence of voice loudness on the same phoneme can be removed.

Note that the present invention is not limited to the above embodiments.

For example, in the previous embodiment, in order to obtain the integrand of equation (5), the spectral envelope Pi) output from the extraction unit 11
Although a case has been described in which the logarithmization is performed and then the normalization is performed, it goes without saying that the reverse process may be performed.

Furthermore, it goes without saying that the speech analysis device of the present invention may be implemented using either hardware or software.

It goes without saying that various other modifications may be made without departing from the gist of the invention.

[Effects of the invention] According to this invention, it is possible to obtain a phoneme-specific analysis pattern without being affected by the length of the vocal tract, so it is possible to obtain the same phoneme regardless of the speaker, such as speech recognition of unspecified speakers. This can be of great help if you want to obtain the same analysis results.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of one embodiment, Fig. 2 is a spectrum diagram for explaining the outline of one embodiment, and Fig. 3 is a block diagram showing the configuration of one embodiment.
4 is an adjustment diagram for explaining the outline of one embodiment,
Fig. 5 is a spectrum envelope diagram for explaining the outline of one embodiment, Fig. 6 is a characteristic diagram for explaining the Mel scale, and Fig. 7 is a spectrum diagram for explaining the outline of one embodiment. 8 to 15 are spectral diagrams for explaining actual measurement examples of one embodiment. DESCRIPTION OF SYMBOLS 11... Spectrum envelope extraction part, 12... Logarithmization part, 13... Normalization part, 74... Integration part, 15...
Projection department. Applicant's agent Patent attorney Takehiko Suzue 0 500 1000 F, (H2
) g3 figure I PI F2 T (possible - acceptable) ÷ (1 - disastrous figure 5 Lp (↑n' Showa residence ° 1277 Japan Patent Office Commissioner Michibe Uga 2, title of invention Speech analysis device 3, amendment Related Patent Applicant (319) Toppan Printing Co., Ltd. 4, Agent 5, Voluntary Amendment 7, Contents of Amendment (1) On page 3, line 4 of the specification, it says ``1 former.'' (2) Similarly, in the 6th line of page 3, the word “latter” should be corrected as “
The former,” he corrected. (3) Similarly, on page 6, line 17, the word ``comrades'' is corrected to ``doujin''.

Claims (1)

[Scope of Claims] A spectral envelope extraction means for extracting a spectral envelope from an input audio signal; a conversion means for converting the spectral envelope extracted by the spectral envelope extraction means into a normalized logarithmic spectral envelope; and this conversion. an integrating means for integrating the normalized logarithmic spectral envelope outputted from the means on a Mel scale; and projecting envelope information of the spectral envelope extracted by the spectral envelope extracting means from the frequency onto the integral output of the integrating means. A speech analysis device comprising a projection means for