JPS61190400A - Enunciation speed estimate apparatus - 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] [Summary] This invention estimates the pronunciation speed of speech, particularly of five vowels and pellicles, and in particular estimates the maximum value at each time based on the similarity time series of the five vowels and pellicles of input speech. The maximum value of the autocorrelation value of the selected time series is extracted, and the reciprocal of the obtained maximum point is taken as the vocalization rate, thereby obtaining the unit time vocalization rate of the five vowels and the pellicles.

[Industrial application field]

The present invention relates to speech processing devices, and particularly relates to a speech rate estimating device for estimating the unit time speech rate of five vowels and pellicles.

[Problems to be solved by conventional technology and invention]

A conventional method for estimating speech rate will be explained using FIGS. 3(a) to 3(e). In FIG. 3(a), the vertical axis represents audio energy (E) and the horizontal axis represents time (1). In this case, for example, when uttering "AKADA J", the energy of the vowel part increases in the vocal section (T), resulting in three peaks. Such a speech waveform is defined as a function E = g (t),
As shown in Figures 3(b) to (e), for example, the correlation function between the audio energy waveform and the sine wave function is calculated by applying a cosine series expansion coefficient based on the number f (cos θ) between 1 to 4 cycles of the sine wave. demand. In other words, the function g(t) and the function r (cos θ
), and the inner product is called the expansion coefficient. In this case, when calculating the inner product of the 1-cycle waveform and g(t) as shown in (b), as is clear from the figure, A
The absolute value of the expansion coefficient becomes small because the waveform of that part is significantly shifted. In this way, for (b) and (e), A
The absolute value of the expansion coefficient becomes small because the waveform of the part shifts. On the other hand, in the case of 3 cycles (d), the waveform of A and the negative peak almost match, so the absolute value becomes large.

In this way, calculate the inner product for the 4th cycle and beyond,
The absolute value is maximum when synchronized in the positive or negative direction with respect to g(t), so this cycle number (frequency) is the number of vowels per second in the speech interval T (Mo 97 seconds). ).

In this way, conventionally, for example, as a minute section of a voice section, the number 1 is
Regarding the time series of energy values every 0 m5ec, the frequency of the cosine expansion function where the absolute value of the cosine series expansion coefficient becomes the maximum was set as the utterance rate m:σ.

However, although conventionally, each mora (five vowels and pellicles) forms a peak in the energy time series in this way, it does not necessarily form a clear peak and valley in actual speech. “AI (AI)J is 2
As the peaks continue due to the two vowels, the sine wave function cannot be applied well, so the frequency cannot be determined and the speech rate may not be extracted accurately.

[Means and effects for solving the problems] The present invention is a speech rate estimating device that solves the above-mentioned problems, and which calculates the maximum value time series of the similarity time series of each mora-forming phoneme (5 vowels, plosives). is a mora position time series, the maximum point of the autocorrelation function is extracted as the mora repetition time, and the reciprocal thereof is taken as the speaking rate. an acousto-electric conversion unit that converts audio waves into electrical signals; an input unit that converts the electrical signals into digital signals; a feature extraction unit that extracts features in time series from the digital signals;
a dictionary section that stores B sound features in advance; a similarity calculation section that performs time-series calculation of similarity between the output of the feature extraction section and the features of each sound read from the dictionary section; an optimal value selection unit that selects the optimal value of the similarity of each sound at each time;
an autocorrelation calculation unit that calculates the autocorrelation of the output time series of the optimal value selection unit over the entire speech interval; a maximum extraction unit that extracts the maximum point of the autocorrelation; and a reciprocal calculation unit that calculates the reciprocal of the maximum point. , and the reciprocal number is taken as the speech rate per unit time.

〔Example〕

FIG. 1 is a block diagram of an embodiment of a speech rate estimating device according to the present invention. In FIG. 1, audio input from a microphone 1 serving as an acoustoelectric transducer is digitized by an input unit 2, and frequency-analyzed by a feature extraction unit 3 every several IQmsec to identify energy characteristics within about 16 bands. Extracted as a vector. The dictionary section 4 stores feature vectors for each phoneme, which are obtained by extracting the features of the five vowels and pellicles in the same manner as described above. The similarity calculation unit 5 calculates the similarity (ρ) between the feature vector (x,) extracted in the feature extraction @3 and the feature vector (y,) for each phoneme in the dictionary unit 4, for example, as shown below ( 1) It is carried out as shown in the formula. Here, ρA is a value indicating the degree of similarity with "A".

Below, the margin feature vector (xi) is a number I as a minute interval of the analysis period.
Since the similarities are obtained in Qmsec degrees, the similarities ρ1 to ρ9 are also calculated every time a feature vector is obtained, and at time j, ρ4. ρj, ρj, ρ4. ρ4. A similarity time series of each phoneme called ρj is obtained.

The optimum value selection unit 6 performs the following calculation.

Here, ff1ax indicates that the maximum value is selected as the optimal value of the argument.

The autocorrelation calculation unit 7 calculates the autocorrelation function of ρ calculated in the previous stage using the following formula.

The maximum extraction unit 8 extracts the maximum peak point of (νk)k-IN other than 0 as □8, and multiplies it by the analysis period T (sec) from the following equation (4) to obtain the average time of 1 mora. gain length.

L = T −k,,, −−−−
--- (4) The reciprocal calculation section 9 calculates the reciprocal of L as shown in the following equation to obtain the speaking speed S (Mo 97 seconds).

S=1/L -.-(5) FIGS. 2(a) to 2(h) are diagrams for explaining an example of the above-mentioned processing procedure. For example, taking "GINZANO J" as an example, in the figure, the vertical axis represents the audio energy value, and the horizontal axis (D represents time (for example, 1Qmsec interval).

In (a), the similarity ρjA at time j is voice rG
It can be seen that the degree of similarity is high in rAJ among INZANOJ. Similarly r I J rUJ---1・-r
It can be seen that where the NJ similarity is high, the waveforms look like mountains (b) to (f). These similarity calculations are performed in the similarity calculation section 5 based on the above-mentioned equation (1).

and j from 1st to n1th, that is, 10m5e
Perform the above similarity calculation by changing each c sequentially. r A J
Find the maximum value for r I J ----------rNJ. This calculation is shown as equation (2) and is calculated in the optimum value selection section 6. The result is a waveform as shown in (g). When an autocorrelation function is calculated based on equation (3) for such a waveform, a waveform shown in (h) is obtained. That is, here, the vertical axis (ν,) shows the autocorrelation function value, the horizontal axis (k) shows the shift amount, and the smoothness lk is (g
) indicates that the waveform shown in ) is shifted only in time. By calculating the autocorrelation function between the waveform (g) shifted only in this way and the waveform (g) at =0, the waveform shown in (h) is obtained. Of course, it is natural that the correlation becomes large when ='O1, that is, when there is no shift, and the next largest point k max is obtained except for k=0. In the reciprocal calculation unit 9, k max at this time is multiplied by the analysis period T as shown in equation (4) to obtain the average time length per mora. By calculating the reciprocal S of this L, the generation rate can be obtained.

〔Effect of the invention〕

According to the present invention, it is possible to accurately estimate the speaking rate of speech without explicitly performing phoneme recognition.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an embodiment of the speech rate estimating device according to the present invention, FIGS. 2(a) to (h) are diagrams explaining the processing procedure of the device in FIG. FIGS. 8A to 8E are diagrams illustrating a conventional method for estimating speech rate. (Explanation of symbols) 1-microphone, 2-human power section, 3...-feature extraction section, 4--dictionary section, 5--similarity calculation section, 6-optimum value selection section, 7.--autocorrelation calculation unit, 8-maximum extraction unit, 9-reciprocal calculation unit.

Claims (1)

[Claims] 1. In a speech rate estimation device in a speech processing device,
an acousto-electric conversion unit that converts a speech wave into an electrical signal; an input unit that converts the electrical signal into a digital signal; a feature extraction unit that extracts features in time series from the digital signal; a similarity calculation unit that performs time-series calculation of similarity between the output of the feature extraction unit and the feature of each sound read from the dictionary unit; an optimal value selection section that selects the optimal value of sound similarity; an autocorrelation calculation section that calculates the autocorrelation of the output time series of the optimal value selection section over the entire speech interval; and an autocorrelation calculation section that calculates the maximum point of the autocorrelation. A speech rate estimating device comprising: a maximum extraction unit that extracts; and a reciprocal calculation unit that calculates a reciprocal of the maximum point; the reciprocal is used as a vocalization rate per unit time.