JPH0145640B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a noise suppression method for a speech signal such as an output signal of a speech synthesis system. As shown in Figure 1, the waveforms of voices, especially vowels such as A, I, and U, are in a specific frequency range of 1K in this example.
Energy is concentrated near Hz, 2KHz, and 3.6KHz. This is due to vocal tract resonance and is called formant, and these first, second, and third formants
The dips d ₁ and d ₂ between F ₁ , F ₂ , and F ₃ are due to antiresonance of the vocal tract and are called antiformants. In this way, in a signal with formant and antiformant, the signal-to-noise ratio S/N of the formant part is sufficiently large, but the S/N of the antiformant part is _small .
(This can be easily understood from the fact that there is not much difference at _d2 , and it is below the noise level at d2), and the S/N deteriorates overall. As a countermeasure to this problem, a method has been considered in which a notch filter that band-eliminates the antiformant part (the shaded part indicates the removed part) is inserted into the output terminal of the decoder of the speech synthesis system (for example, TSLamba etal: Intelligible
Voice Communication……，Radio and
Electronic Engineer, Vol.48, No.4). In this case, not only the noise component but also the signal component in the relevant band is lost, but the formant component of the original signal is preserved, so there is little deterioration in voice quality and the noise characteristics can be improved (it sounds like there is no noise). However, the formant frequency changes significantly depending on the type of voice, such as the type of vowel, as shown in the table below, so a notch filter that is effective for some types of voice will actually degrade the quality for other types of voice. I will do it.

【table】

[Table] The above drawbacks were audibly confirmed when the center frequency and band of the notch filter were changed when various sounds were written to memory and read out and decoded using a system similar to the system described in the above literature. . In order to eliminate such disturbances, the present invention provides a notch filter that can automatically change the center frequency and band, and changes the center frequency and band according to each voice, thereby eliminating noise while maintaining the fidelity of the voice. This is an attempt to improve the characteristics. Next, this will be explained in detail with reference to examples. FIG. 2 shows an example in which the present invention is applied to a waveform encoding type speech synthesis system. 10 is the storage section,
It is composed of a ROM (read-only memory) and has a part A for storing waveform encoded speech segments and a part B for storing characteristic control information of a notch filter used for the speech segments. In the drawings, parts A and B are shown so that one is separated from the first half of the memory and the other is separated from the latter half of the memory, but this is for illustrative purposes only, and the actual arrangement on the memory system can be changed as appropriate. Reference numeral 12 denotes a control section composed of a microprocessor, which outputs a read address signal for the storage section 10 and the like. A decoder section 14 decodes the read digital output of the storage section 10 and outputs an audio analog signal. 16 is a notch filter section, and the center frequency and width of the band to be removed are changed by the control information. For example, part A of the storage section
If a vowel having the frequency spectrum characteristics as shown in FIG. 1 is to be read out from a group of speech segments stored in The address group of the part A that stores the digital data group is sequentially outputted, and the read output is decoded by the decoder section 14 to become an analog audio signal, which is outputted through the notch filter section 16 and then sent through an amplifier (not shown). The notch filter unit 16 regarding the vowel stored in part B, which is the first address of the address group output by the control unit, drives the speaker.
Characteristic control information, ie, a control signal for setting the center frequency and band of the filter to the center and width of the hatched area in FIG. 1 in this example, is also read out, and the characteristics of the notch filter section 16 are changed as described above. Since the information written to the memory is digital information, if the filter control signal is an analog signal, DA conversion is performed. Note that a digital filter is preferable as this notch filter. Further, a plurality of notch filters may be provided to remove the first, second, . In general speech, the silent period is 50%, the vowel interval is 49%,
There is also data published in the literature that the consonant period is 1%, and the characteristic control of the notch filter is useful when distinguishing speech from syllables, phonemes, vowels, consonants, etc.
It is effective to do this on a phoneme or vowel basis.
It is also effective to provide this filter section with a function of expanding the formant section. Note that the audio signal is
When writing a group of speech segments into the storage section A by modulation, etc., it is also possible to pass the speech signal through a notch filter to remove the antiformant part. The method of the present invention is effective for speech synthesis output using a so-called general-purpose speech memory in which a group of speech segments is written without performing antiformant part removal. As explained above, according to the present invention, meaningful speech with good sound quality and excellent noise characteristics is output using a group of speech segments such as A, I..., Ka, Ki... stored in the memory. It can be done and is very effective.

[Brief explanation of drawings]

FIG. 1 is a graph showing the frequency spectrum density characteristics of a speech segment, and FIG. 2 is a block diagram for explaining the present invention. In the drawing, 10 is a memory, 14 is a decoding section, 16
is a notsuchi filter.

Claims (1)

[Claims] 1. In a noise suppression method for a system that selectively reads out a large number of speech segments stored in a memory, decodes them, and outputs a speech signal, a notch filter for removing the antiformant part of each speech segment is provided, and its characteristic control information is provided. is stored in memory in correspondence with each speech segment, and when each speech segment is read out from the memory, characteristic control information for the speech segment is also read out and added to the notch filter. A method for suppressing noise in a speech signal, characterized in that a decoded output of a speech segment readout output is passed through a notch filter.