JPS5848917B2 - Smoothing method for audio spectrum change rate - Google Patents

Abstract

PURPOSE:To enable to grasp the change in spectrum at tone boundary clearly, by eliminating the fluctuation in timely spectrum change, through weighting the difference of envelope of audio spectrum.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for smoothing a speech spectrum change rate used in various speech analyzes.

In speech analysis, it is well known that it is very effective to use the spectral change rate when dividing natural speech into individual phonemes (called segmentation).

Although the speech spectrum is stable in the phonological stationary region,
Phonological boundaries change significantly over time.

Therefore, phoneme divisions can be determined based on the temporal characteristics of spectral changes.

As for this spectrum change rate, the fluctuation is small,
It is also desirable to have something that can accurately capture sudden changes.

As a calculation method, a method of determining a change in the spectrum between two adjacent points of time, a method of determining a time series of the difference from a standard spectrum prepared in advance, etc. are often used.

However, the values obtained by these calculation methods generally have large fluctuations, making it difficult to use them as they are as parameters for segmentation, and some kind of smoothing processing is required.

Conventional smoothing methods include storing the spectral change rate calculated for each frame in a buffer memory, using a delay element and taking the arithmetic average of the previous one in time, and using a low-pass filter. However, when these methods are used, they not only smooth the spectral change rate, but also distort the steep changes, resulting in a loss of accuracy.

The present invention has devised a smoothing method that removes fluctuations in the rate of spectral change and does not lose its steep changes.

Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 shows an embodiment of the method for smoothing the spectral change rate according to the present invention.

First, an audio signal is input from a terminal 1, and is subjected to low-pass filtering at 4 KHz in an audio signal preprocessing section 2, sampled at 8 KHz, and further quantized at 11 bits to be converted into a digital audio signal.

The spectral envelope calculation section 3 calculates the spectral envelope from this digital audio signal using FFT (Fast Fourier Transform) or the like, and stores it in the buffer memory 4.

In calculating the spectral change rate in this embodiment, the Mel conversion circuit 5 performs Men conversion on the frequency axis in order to reflect the auditory characteristics.

This conversion is performed according to the following equation. Mel(f)'=C
-log(1000+f) (C: constant, f: frequency) Using the Mel-transformed spectrum envelope, the spectrum envelope difference calculating circuit 6 calculates the difference between the spectrum envelope and the spectrum envelope separated by τ (msec.).

This envelope difference d(t, t-τ) is calculated by the following equation.

4000 (Hz) d(t, t-r) 2f lA(r, t)
-a(V, t-τ) l”dMe6(f)300(H
, 9 Here, A(f, t) is the power (dB) of the frequency component f of the spectrum envelope at time t.

This smoothing method uses d(t, t-
Instead of considering τ) as the spectral change rate at time t and smoothing it, −
30<τ<+30 (msec.),
The products obtained by multiplying d(t, t-τ) by the weight W(τ) are added together, thereby achieving smoothing.

Spectral change rate: Smoothing is performed by looking at spectral changes over intervals in this way, so the spectral change rate obtained by the above process not only has small fluctuations but also accurately represents steep points of change. .

For these processes, a coefficient multiplier 7 and an integrator 8 shown in FIG. 1 are used.

In the coefficient unit 7, the envelope difference d(t, t-τ) obtained by the spectral envelope difference calculation circuit 6 is multiplied by the weighting function W(τ).

These are sequentially added by integrator 8, and the result is output at terminal 9.
is output from.

An example showing the above smoothing effect is shown in FIG.

The upper part of Figure 2 shows the power of the voice uttered "Blue Airship", and the middle part shows the power of the voice uttered "Blue Airship", and the middle part shows the power of τ = 10 msec. It shows the difference d(t, t-10) from the distance,
This is the output directly from the spectral envelope difference calculation circuit 6 to the terminal 9 except for points 7 and 8 in FIG. 1 (indicated by dotted lines in FIG. 1).

The lower part shows this smoothing method.

It can be seen that the effect is clearly expressed compared to the one in the middle row.

The above embodiment} Here, the various constants, weighting functions,
Spectrum calculation method (well, it is not limited to these, but the important thing about this smoothing method is to set up a window that includes the frequency for which the spectral change rate is to be calculated, and then take a weighted average of the spectral change rate over this window) This is to perform smoothing.

As explained above, the method for smoothing the rate of change in the audio spectrum of the present invention does not simply obtain the change in the spectrum between two different points in time, but rather calculates the weighted average of the difference in the spectrum within a window over a certain period of time. Since this is a method for calculating the amount of spectral change, it is possible not only to remove temporal fluctuations in spectral change and smooth it, but also to weight the spectral difference between adjacent points of time to a greater extent, so that the spectral change at the phonetic boundary can be calculated. It has the advantage of being clear.

[Brief explanation of the drawing]

FIG. 1 is an example of the method for smoothing the rate of change in the audio spectrum of the present invention. FIG. 2 shows the power in the upper row and the rate of spectral change before smoothing in the middle row for the voice uttered "Blue Airship." The lower part is a diagram showing the spectral change rate after processing. DESCRIPTION OF SYMBOLS 1...Input terminal, 2...Audio signal preprocessing section, 3...Spectral envelope calculation section, 4...
... buffer memory, 5 ... Men conversion circuit, 6 ... spectrum envelope difference calculation circuit, 7 ...
... Coefficient unit, 8 ... Integrator, 9 ...
・Output terminal.

Claims (1)

[Claims] 1. In a method for calculating the rate of spectral change using various parameters that characterize the speech spectrum, a window is provided that includes the frequency for which the rate of spectral change is to be determined, and the weighted average of the spectral differences within this window is taken to calculate the rate of spectral change. A method for smoothing a rate of change in an audio spectrum, characterized by performing smoothing.