JPH0358520B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (a) Field of Industrial Application The present invention relates to a speech synthesis method for synthesizing speech based on speech characteristic parameters. (b) Prior art This type of speech synthesis method electrically simulates the functions of the human generator, starting with the conventional vocoder technology, LPC (Linear Predictive Coding) method,
The PARCOR (partial autocorrelation) method and the LSP (line spectral pair) method have been developed one after another. Among them, the recently developed LSP method has an extremely high information compression ratio of LPS frequency parameters used as voice characteristic parameters, and has the advantage of being able to obtain high-quality synthesized speech. Speech synthesis methods are becoming mainstream, replacing synthesis methods. Figure 1 shows the basic configuration of a speech synthesizer that uses the LSP method. In the figure, 1 is a voiced sound source generation circuit that simulates the periodicity of human vocal cord vibration to generate a periodic pulse sound source signal, and 2 is a noise source that simulates the aperiodicity of turbulent vibration in the human trachea. This is a silent sound source generation circuit that generates a signal. Reference numeral 3 designates a selection switch for selecting the periodic pulse sound source signal obtained from the voiced sound source generation circuit 1 or the noise sound source signal obtained from the unvoiced sound source generation circuit 2. 4 is a multiplier that adds the original amplitude component of the signal to the sound source signal obtained from the switch 3; 5 is a digital filter that simulates the acoustic characteristics of the human vocal tract; By filtering the unvoiced sound source signal, a voice waveform signal of voiced speech or unvoiced speech is output. 6 is a D/A converter that converts the digital audio waveform signal obtained from the digital filter 5 into an analog value; 7 is the D/A converter 6;
This is a speaker that emits synthesized speech based on the audio waveform signal from. Reference numeral 8 denotes a parameter memory, which sets the period of the periodic pulse sound source signal in the voiced sound source generating circuit 1 and also controls the selection switch 3.
a pitch parameter P that instructs the selection of the sound source signal at , an amplifier parameter A that sets the amplitude component in the multiplier 4, an LSP frequency parameter f ₁ that determines the filter characteristics of the digital filter 5,
f ₂ , ..., f ₈ are stored for each frame period. Therefore, when synthesizing unvoiced speech, the value of the pitch parameter P is "0", and the selection switch 3 detects this value "0" and selects the noise source signal from the unvoiced sound source generation circuit 2. By introducing the signal into the multiplier 4 and multiplying this sound source signal by the value of the amplifier parameter A, an unvoiced sound source signal is obtained. On the other hand, when synthesizing voiced speech, the value of the pitch parameter P indicates the pitch period depending on a value other than "0", and depending on this value P, the sound source generating circuit 1 determines the period of the periodic pulse sound source signal. At the same time, the selection switch 3 detects that the value of the pitch parameter P is not "0" and introduces the periodic pulse sound source signal from the voiced sound source generating circuit 1 into the multiplier 4, and adds the value of the amplifier parameter A to this sound source signal. By multiplying , a voiced sound source signal can be obtained. Thus, the obtained voiced or unvoiced source signal is LSP
It is filtered by a digital filter 5 whose filter characteristics are controlled by parameters f ₁ , f ₂ , ..., f ₈ , and further subjected to D/A conversion 6 and then voiced or unvoiced synthesized speech is uttered by a speaker 7 . In such a speech synthesizer, the parameters P, A, which should be stored in the parameter memory 8 are
It is necessary to analyze f in advance based on the original audio signal, and the analysis process of LSP frequencies f ₁ , f ₂ , ..., f ₈ is described in "Nikkei Electronics" published on February 2, 1981. For details, see the article "Speech synthesis method using line spectrum frequency as a parameter and its LSI implementation" on pages 128 to 158. Furthermore, for analysis of pitch parameter P and amplifier parameter A, the analysis process used in the conventional PARCOR method is used. In particular, the condition for determining voiced/unvoiced depending on whether or not the value of the pitch parameter P is set to "0" is described in detail in Japanese Patent Publication No. 55-34953, using the modified correlation function W(τ ) is detected, and the first-order
Find the value of PARCOR coefficient K1 multiplied by 0.5 and add 0.5K1, that is, Pm+0.5K1, and compare this value with a specific threshold t. When Pm+0.5K1≧t, the original audio signal is voiced speech. On the other hand, when Pm+0.5K1<t, the original audio signal was determined to be unvoiced audio. However, in the LSP speech synthesis method, when determining voiced or unvoiced in speech analysis,
Not only does it require the use of the PARCOR coefficient K1, but it also has the disadvantage of requiring a large amount of calculation processing to derive the maximum value Pm of the variable autocorrelation function W(τ).
There was an inconvenience that hindered real-time processing of such speech analysis. (c) Object of the Invention The present invention has been made in view of the above points and provides a speech synthesis method that simplifies the voiced/unvoiced determination process. (d) Structure of the invention The speech synthesis method of the invention has LSP parameters f ₁ ,
The average values _{1 , 2} , ... and standard deviations σ ₁ , σ ₂ , ...σ _N of f 2 , ..., f _N are determined in advance, and
After normalizing the LSP parameters f ₁ (T), f ₂ (T), ..., f _N (T) by respectively converting them into Fi (T) = fi (T) - fi/σi, further weighting is performed. The coefficients C ₁ , C ₂ , ..., _CN are set to calculate A(T)= _N 〓 ⁱ⁼¹ CiFi(T), and voiced/unvoiced is determined based on this value A(T). be. (E) Embodiment FIG. 2 shows an embodiment of the original speech analysis device in the speech synthesis method of the present invention. Reference numeral 10 denotes a microphone for inputting the original audio, and 11 an A/D converter for converting the audio signal from the microphone 1 into a digital format. 12 is the A/D converter 11
This is an LSP analysis circuit that LSP-analyzes the audio signal from . Basically, it has a filter characteristic opposite to that of the digital filter 5 of the synthesizer shown in _FIG . f ₂ (T),
..., f ₈ (T) is calculated, and the amplifier parameter A(T) is output as the energy value of the residual signal corresponding to the sound source signal. 13 is the above A/D
This is a pitch extraction circuit equipped with a peak hold function to hold the peak value of the audio signal from the converter 11, and outputs the time between peaks as a pitch period P'. 14 is a storage unit, in which each LSP parameter f ₁ ,
Regarding f ₂ , ..., f ₈ , the existence distribution of each value is determined in advance, and the average value ₁ , ₂ , ..., ₈ of each value and its standard deviation σ ₁ , σ ₂ , ..., σ ₈ are calculated in advance. Each value is calculated and stored. 15, 15... are the first to eighth normalization circuits, and the i-th (i=1, 2,..., 8)
The normalization circuit 15 converts the i-th LSP frequency parameter fi(T) obtained at this time from the LSP analysis circuit 12 into the average value and standard deviation σi of the i-th LSP frequency parameter in the storage unit 14. is used to convert it into a value Fi(T) expressed as Fi(T)=fi(T)-fi/σi. That is, this value Fi
(T) is a normalized representation of how much each LSP frequency parameter fi(T) deviates from the average value in this frame period T. 1
6, 16... are the first to eighth multipliers, and the i-th multiplier
The value Fi(T) from the th normalization circuit 15 is multiplied by the ith weighting coefficient Ci to calculate the value CiFi(T). 17 is an adder, which calculates the sum of the values CiFi(T) from the first to eighth multipliers 16, 16, . . . A(T)= ₈ 〓 ⁱ⁼¹ CiFi(T). That is, when this value A(T) is large, each LSP frequency parameter fi(T) at this frame period T deviates greatly from the standard value,
Conversely, when this value A(T) is small, it indicates that it is close to a standard value. 1
8 is a comparison circuit that compares the value A(T) obtained from the adder 17 with a specific threshold value Vth, and when A(T)>Vth, each LSP frequency parameter fi
(T) is determined to be significantly deviated from the standard value, that is, it is determined to be unvoiced voice, and an unvoiced voice detection signal is output. Conversely, when A(T)≦Vth, each LSP frequency parameter fi
(T) is determined to be within a standard value range, that is, it is determined to be voiced speech, and a voiced speech detection signal is output. 19 is the pitch period P' from the pitch extraction circuit 13 and the value "0" from the zero register 20.
This is a changeover switch that outputs the pitch period P' when it receives the voiced voice detection signal from the comparison circuit 18, and outputs the value "0" when it receives the unvoiced voice detection signal. It is structured as follows. That is, the pitch parameter P is output from this changeover switch 19. Each parameter obtained in this way f ₁ , f ₂ , ..., f ₈ ,
If A and P are stored in the parameter memory 8 in FIG. The sound source signal or the periodic pulse sound source signal from the voiced sound source generation circuit 1 is selected. Moreover, the parameter f ₁ obtained as described above,
If f ₂ , ..., f ₈ , A, and P are directly transmitted to the digital filter 5, multiplier 4, voiced sound source generation circuit 1, and selection switch 3 without storing them in the parameter memory 8 of FIG. Analysis and synthesis of time becomes possible. Next, taking as an example the analysis of the voice "I will inform you", a specific example of the voiced/unvoiced judgment condition will be shown with reference to FIG. LSP frequency parameters f ₁ , f ₂ ,…,
The order of f _N is N=8, and each weighting coefficient is all
Set Ci=1 (i=1, 2, ..., 8), and further set the threshold
When Vth=N/2=4, as shown in Figure 3, A(T)>Vth is `` OSH IRA SE SH IMA
Among the sounds of "SU", two time regions T _a and T _c correspond to "SH" which is an unvoiced sound, and two time regions Tb and Td correspond to "S" which is also an unvoiced sound. Conversely, the time domain other than these is the time domain of voiced speech. Also, these time domain Ta,
Tb, Tc, and Td are approximately the same as the unvoiced speech judgment regions Ta′, Tb′, Tc′, and Td′ using the peak value Pm of the modified correlation function and the first-order RARCOR coefficient K1 in the conventional method. I understand. In the above embodiment, the average values ₁ , ₂ , ..., ₈ of the LSP frequency parameters stored in the storage unit 14 of the speech analysis device and their standard deviations σ ₁ , σ ₂ ,
..., σ ₈ is the LSP frequency parameters f ₁ , f ₂ , ..., extracted for all voiced and unvoiced voices.
You may use the one derived based on f ₈ , but
If we derive this based on the parameters f ₁ , f ₂ , ..., f ₈ extracted for all voiced voices, we get
Not only can the calculation processing for this purpose be reduced, but also more accurate voiced/unvoiced determination processing can be expected. (F) Effects of the Invention As is clear from the above explanation, the speech synthesis method of the present invention adopts the LSP method and uses the value A(T) formed by the LSP frequency parameter to determine voiced or unvoiced speech. Therefore, compared to the conventional method using the maximum value of the variable autocorrelation function, which requires a large amount of calculation processing, the amount of calculation for the voiced/unvoiced determination processing can be significantly reduced. Therefore, the voiced/unvoiced determination processing can be simplified, and real-time speech analysis processing can be expected.

[Brief explanation of drawings]

Fig. 1 is a basic configuration diagram of a speech synthesis device employing the LSP method, Fig. 2 is a configuration diagram of a speech analysis device used in the speech synthesis method of the present invention, and Fig. 3 is a diagram of the value A ( It is a time curve diagram of T). 1... Voiced sound source generating circuit, 2... Unvoiced sound source generating circuit, 3... Selection switch, 5... Digital filter, 8... Parameter memory, 12... LSP
Analysis circuit, 14... Storage unit, 15... Normalization circuit, 16... Multiplier, 17... Adder, 18...
Comparison circuit, 19...changeover switch.

Claims (1)

[Scope of Claims] 1. A voiced sound source generation circuit that generates a periodic pulse sound source signal for voiced speech, and an unvoiced sound source generation circuit that generates a noise source signal for unvoiced speech; and a digital filter that filters the sound source signal obtained from one of the circuits.
In a speech synthesis method that synthesizes speech waveforms by controlling the filtering characteristics of the digital filter using N line spectrum pair parameters f ₁ , f ₂ , ..., f _N , the line spectrum pair parameter f ₁ , _f2 ,...
Each mean value of f _{N 1} , ₂ , ..., f _N and standard deviation σ ₁ , σ ₂ ,
..., σ _N are calculated in advance, and the line spectrum pair parameters f ₁ (T), f ₂ (T), ..., f _N (T) at time T are respectively Fi (T) = fi (T) - After normalization by converting to fi/σi (i=1, 2,...N), weighting coefficients C ₁ , C ₂ ,..., C _N are further set to obtain A(T)= _N 〓 ^{i =1} CiFi(T) (i=1, 2,...N) is calculated, and when this value A(T) is greater than a specific threshold value Vth, the noise source signal from the unvoiced sound source generation circuit is A voice synthesis apparatus characterized in that a periodic pulse sound source signal from the voiced sound source generation circuit is supplied to the digital filter when the value A(T) is smaller than a specific threshold value Vth. method.