JPS5912198B2 - Audio parameter abnormality detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Technical Field of the Invention The present invention is directed to a method for analyzing and synthesizing speech that detects abnormal speech parameters that can cause difficult-to-hear sounds such as strange noises to be generated in synthesized speech. The present invention relates to a method for automatically detecting an abnormal parameter, and in particular to a method for detecting an abnormal parameter by approximately converting it into a function of polar frequency and bandwidth.

(2) Technical background Generally, in linear prediction methods such as PARCOR and LSP, an all-pole model is represented by H(z)=1+ΣαiZ−1 i=1, where P is the prediction order and αi is the linear prediction coefficient. This model approximates the spectral envelope characteristics of speech, so this model often does not apply to nasal consonants such as N, M, 10, and the like, or the vowel I, which has a low first formant frequency. In the case of voice and the pitch frequency almost matching the first formant frequency, the bandwidth of the first formant may be analyzed as being abnormally narrow.

In such cases, the synthesized speech often has a very large amplitude or contains abnormal sounds that are difficult to hear. (3) Prior art and problems Conventionally, when analyzing and synthesizing speech, certain I had to repeat the process of searching for the point in time where the parameters were, and then making appropriate modifications to the parameters.

Moreover, in the past, it was not always possible to accurately determine which part of the voice parameters caused an abnormality in a synthesized voice. In order to suppress this problem, a trial-and-error and inefficient correction process was required. 30(4) Object of the Invention The object of the present invention is to provide a means for automatically detecting an abnormality in a parameter that is particularly likely to cause an abnormal sound.

(5) Structure of the Invention 35 The present invention automatically detects abnormal parameters, so that the proximity of the pitch frequency and the polar frequency with a narrow bandwidth in the spectrum envelope is This was done with a focus on the fact that it is closely related to development.

FIG. 1 is an explanatory diagram of the relationship between the above-mentioned pole frequency and pitch frequency.

In the figure, F is a plurality of polar frequencies Fi and a bandwidth Bi
This is the spectral envelope approximately expressed by In the present invention, among the bandwidth Bi of each pole frequency, a pole frequency Fi (for example, 300 Hz) having a particularly narrow bandwidth Bi,
If the pitch frequencies Pi are close to each other with a difference of, for example, several Hz to 30 Hz, this is determined to be an abnormal parameter. The present invention provides a speech analysis and synthesis method in which speech is analyzed to extract information on pitch frequencies and information on spectral envelope, and this is used as a parameter time series for speech synthesis. means for approximately converting the polar frequency into a function of a polar frequency and its bandwidth; means for extracting a polar frequency having a narrow bandwidth for the transformed function; The present invention is characterized in that it includes means for determining the degree of proximity between the polar frequency and the pitch frequency, and when the degree of proximity between the polar frequency and the pitch frequency is high, it is determined that this is a parameter abnormality.

(6) Examples of the invention The present invention will be described in detail below using examples.

FIG. 2 is a block diagram of a speech analyzer embodying the present invention.
In the figure, 1 is a voice analysis processing section, and 2 is a voice parameter abnormality detection section according to the present invention. In the speech analysis processing section 1, 3 is a sound source information analysis section that extracts pitch, amplitude, voiced/unvoiced information from input speech, and 4 is a linear predictive analysis and PAR for parameterizing spectral envelope information.
This is a linear predictive analysis unit that performs COR conversion. Analysis department 3,
The analysis output from 4 is used for speech synthesis in a speech synthesizer, not shown. In the parameter abnormality detection unit 2, 5 is the PARCOR output from the linear prediction analysis unit 4.
This is a conversion unit that converts coefficients back into linear prediction coefficients.

If linear prediction coefficients can be extracted from the linear prediction analysis unit 4 as intermediate processing data, the conversion unit 5 is not necessary. 6 is the transfer function H(z) of the all-pole approximation model regarding the spectral envelope of the voice described above, and the equation when the denominator is A(z) is calculated by using the Newton-Raphson method etc. ) This is a root-finding calculation unit that finds roots on the z-plane that satisfy −0.

Reference numeral 7 denotes a conversion unit that calculates the polar frequency and bandwidth, where the solution of the calculation result of the root-finding calculation unit 6 is taken as the sampling period and T is the sampling period.

8 is a part for extracting the polar frequency Fi having a narrow bandwidth Bi.

The degree of narrowness of the bandwidth is determined and extracted depending on whether the value of Bi is less than or equal to a certain threshold value or greater than or equal to a certain threshold value.

Reference numeral 9 denotes a determination unit that determines the degree of proximity between the extracted polar frequency of the narrow bandwidth and the pitch information (frequency) Pi being output from the sound source information analysis unit 3.

The determination of proximity is performed, for example, based on the following equation.

When a polar frequency Fi that satisfies this determination formula is found, it is determined to be an abnormal parameter and output. When the abnormal parameter judgment output is obtained, the data can be used to pre-process the audio analysis output by widening the narrow bandwidth, shifting the pitch frequency, or changing other appropriate factors that affect sound quality. It is possible to correct it.

(7) Effects of the Invention Since the present invention can automatically detect parameters that are likely to cause abnormal speech, it can extremely facilitate the task of improving the quality of synthesized speech.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram of the principle of the present invention, and FIG. 2 is a configuration diagram of an embodiment of the present invention. In the figure, 1 is a speech analysis processing section, 2 is a speech parameter abnormality detection section, 3 is a sound source information analysis section, 4 is a linear prediction analysis section, 5 is a conversion section from PARCOR coefficients to linear prediction coefficients,
Reference numeral 6 indicates a root calculation unit, 7 a conversion unit from roots to pole frequencies and bandwidths, 8 a narrow bandwidth extraction unit, and 9 a proximity determination unit between pole frequencies and pitch frequencies.

Claims (1)

[Claims] 1 In a speech analysis and synthesis method that analyzes speech and extracts information on pitch frequency and information on spectral envelope, and uses this as a parameter time series for speech synthesis,
means for approximately converting the information regarding the analyzed spectral envelope into a function of a polar frequency and its bandwidth; means for extracting a polar frequency having a narrow bandwidth with respect to the transformed function; means for determining the proximity between the analyzed polar frequency and the analyzed pitch frequency,
An audio parameter abnormality detection method that determines a parameter abnormality when the degree of proximity between the polar frequency and the pitch frequency is high.