JPS63121899A - Voice improvement - 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] [Industrial Application Field] The present invention improves the voice of a speaker whose vocal cords do not function properly due to illness or other reasons - so-called vocalization - and adds pitch information. This makes it possible to bring the sound closer to normally generated sounds, thereby making it possible to qualitatively transform the sound.

[Prior Art] In the field of medicine, speech assisting devices such as artificial larynxes have been developed for people who have lost their larynx for some reason. In addition, there are few countermeasures for the voices of people who have vocal cord abnormalities that have not yet been removed, except for vibration devices that apply mechanical vibrations to the vocal cords from the outside. Moreover, in this device, the vibration period (pitch height) is constant, and the intonation cannot be changed depending on the content of the utterance.

[Problems to be Solved by the Invention] Hitherto, there has been no attempt to improve vocalization in a purely engineering manner by artificially processing uttered voices without imposing restrictions on the speaker. The present invention makes this improvement possible to some extent, and aims to make the voice of a speaker with abnormal vocal fold vibrations closer to the voice of a normal speaker through computer processing.

[Means for Solving the Problems] In order to achieve such an object, the hoarseness improvement method of the present invention extracts spectrum information from the voiced part of a hoarse sound signal, configures a vocal tract filter, and improves the vocal tract. It is characterized by manually inputting pitch information extracted from audio signals other than hoarseness signals into a filter to synthesize speech.

[Function] According to the present invention, insufficient pitch information in hoarse voices is extracted from normal speech or artificial vocal fold waveforms, and synthesized with spectral information extracted from vocalizations.
The noise can be improved.

[Example] The present invention will be described in detail below with reference to the drawings.

FIG. 1 shows a block diagram of the youngest son's method. The entire system consists of three parts: an analysis section 1, a feature extraction section 2, and a synthesis section 3, all of which are executed by digital signal processing within an electronic computer. The basis of the present invention is to improve hoarse speech.
In addition to the voice to be improved, information on normal voice with the same content is used, so as shown in the figure, the youngest son's ceremony requires two people in principle to produce the voice and the normal voice. However, if normal speech with the same content cannot be obtained, as a second best measure, an artificial vocal fold waveform can be created and supplied to the synthesis section.

The analysis unit 1 performs analysis (1) for vocalizations and analysis (2) for normal voices.

First, analysis (1) of voice will be explained.

The hoarse sound signal sampled by the D/D converter is divided into a voiced section and a silent section, and the voiced section is further distinguished into an unvoiced sound section and a voiced section. Unvoiced sound sections are recorded as they are, and so-called linear prediction analysis is performed on voiced sound sections to obtain linear prediction coefficients carrying spectral information and residual signals carrying pitch information. At this time, in the analysis (1), that is, the analysis of the voice, the frame length (analysis window width) is fixed (20 to 30 milliseconds) and the frame period is 172 of the frame length, as shown in FIG.

Normal speech consists of a fundamental wave and its harmonics, and has clear pitch information, so it is possible to analyze it in synchronization with it. However, since pitch information is incomplete in speech, the frame length cannot be fixed in this way. , the frame period is shifted and analyzed.

The next feature extraction section extracts only the spectral counts (prediction coefficients) obtained as a result of linear prediction analysis using the frame length and frame period shown in FIG. 2, and discards pitch information (residual signal).

Next, analysis (2), that is, analysis of normal speech having the same content as the speech is performed. In this case, the voiced part is first subjected to pitch synchronization analysis before being subjected to normal analysis. That is, as shown in FIG. 3, the length of the pitch is detected and that section is taken as the frame length. In this case, frames will not overlap. Next, the length of the voiced part of the normal sound is compared with the length of the voiced part of the corresponding utterance, and the length of the normal sound is expanded or contracted to adjust the lengths to be equal with an accuracy of one pitch. FIG. 4 shows details of the analysis (2) part of FIG. 1.

Adjusting the length after performing pitch synchronization analysis is
It has the advantage that expansion and contraction can be performed in units of one pitch, and that expansion and contraction can be averaged over the entire voiced sound section, rather than expansion and contraction of a specific part. In other words, as shown in Figure 5, when the voiced part of normal speech is shorter than the utterance, several pitch sections are inserted at equal intervals, and when it is long, several pitch sections are inserted at equal intervals. The pitch sections are spaced at equal intervals to make the length the same.

In this way, the length of the normal voice is adjusted for each voiced section, and the length of the normal voice is rearranged so that it is almost the same as the corresponding voiced section. After that, analysis of vocalizations (analysis of Figure 1)
The same analysis as in 1)) is performed to obtain spectral information (prediction coefficients) and pitch information (residual signal), but the spectral information is unnecessary and is therefore discarded. In the synthesis section shown in Figure 1, a vocal tract filter is constructed using the spectrum information obtained from the ginsei, that is, the linear prediction coefficients, and the pitch information obtained from the normal speech, that is, the residual signal, is manually applied to the filter. Obtain synthetic speech by doing this. Incidentally, in FIG. 1, when the pitch information is set to be that of a voice, the original sung voice is restored as the output sound, and conversely, when the spectrum information is set to that of a normal voice, the original normal sound is restored. Ginsei originally does not have pitch information, or even if it does, it is extremely incomplete, but during synthesis, by replacing the residual signal of a normal speaker in this way, pitch information can be added, and normal speech can be reproduced. You can get a voice with a tone similar to that of

However, it is not always possible to obtain normal audio with the same content. In such a case, as shown in FIG. 6, an artificial vocal cord waveform simulating a triangular wave is created, and the pitch information obtained from it is supplied to the vocal tract filter. Since the period T of the triangular wave provides pitch information, an average value determined by the gender and age of the speaker is set as the value of T. Further, it is appropriate that the rise (T1) and fall time (T2) of the triangular wave are each several milliseconds. However, in this case, the obtained synthesized sound output uses artificial vocal cord waves, so it becomes a mechanical sounding sound, and the synthesized sound has a sound that is quite different from the real voice.

FIG. 7 shows an example of waveforms of voice and voice improved by the two methods described above. In the same figure, (A) is Ginsei, (B) is the speech waveform improved using normal speech, and (C) is the speech waveform improved using normal speech.
is an improved audio waveform using a triangular wave. Note that when the vocal tract characteristics are changed in this manner, the waveform is slightly deformed, so that discontinuity may occur at the joint of frames during synthesis.

Therefore, as shown in Figure 8, a triangular window with an amplitude of 1 is applied to the synthesized sound output for each frame, and the sum of the gains in the overlapping parts of adjacent frames is always set to 1, so that the waveforms of both It is a good idea to add

Such operations are effective in maintaining waveform continuity and obtaining smooth synthesized sounds.

[Effects of the Invention] As explained above, according to the present invention, insufficient pitch information in hoarseness is extracted from normal speech or artificial vocal fold waveforms, and synthesized with spectral information extracted from the vocalizations. By doing so, the sound can be improved.

[Brief explanation of the drawing]

Figure 1 is a block diagram of the present invention. Figure 2 is a waveform diagram explaining the frame and analysis window width in speech analysis. Figure 3 is a waveform diagram explaining pitch synchronization analysis for normal speech. Figure 4 is a waveform diagram explaining the pitch synchronization analysis for normal speech. A block diagram of the analysis for speech; Fig. 5 is a diagram showing a method for adjusting the length of voiced sections of normal speech; Fig. 6 is an artificial vocal fold waveform diagram; Fig. 7 (A) is the voice of vocalization. Waveform diagram: (B) is a wailing waveform diagram improved using normal speech, (C) is a wailing waveform diagram improved using a triangular wave, and Figure 8 preserves the continuity of the composite waveform. FIG. 1... Analysis section, 2... Feature extraction section, 3... Synthesis section.

Claims (1)

[Claims] 1) A vocal tract filter is configured by extracting spectrum information from a voiced part of a hoarse sound signal, and pitch information extracted from a voice signal other than the hoarse sound signal is input to the vocal tract filter. A method for improving hoarseness, which is characterized by synthesizing speech. 2) The method for improving hoarseness according to claim 1, wherein the pitch information is extracted from a voiced part of a normal speech signal having the same content as the hoarseness signal. 3) The pitch information is extracted after matching the length of the voiced part of the normal speech signal with the length of the voiced part of the hoarse sound signal.
How to improve hoarseness as described in section. 4) The hoarseness improvement method according to claim 1, characterized in that the pitch information is extracted from a triangular wave that simulates a vocal cord waveform. 5) The hoarseness improvement method according to claim 4, characterized in that the period of the triangular wave is set to an average pitch determined by the gender and age of the person producing the hoarseness.