JP2615991B2 - Linear predictive speech analysis and synthesis device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a linear predictive speech analysis / synthesis apparatus.

[Conventional technology]

In a conventional linear prediction speech analysis / synthesis apparatus, when an input speech signal is a voiced sound, an impulse train having a fundamental frequency of the input speech signal as a repetition frequency is generally used as a speech signal on the synthesis side. Further, instead of an impulse train, a pulse train whose envelope waveform has a “shape” that repeats at a fundamental frequency has been used.

[Problems to be solved by the invention]

The above-described conventional linear predictive speech analysis / synthesis apparatus has a drawback that when an impulse train is used as a sound source signal, energy concentrates on a pitch excitation point on a time axis, so that an output speech signal becomes unnatural. When a pulse train having "" is used, concentration of energy can be avoided, but since the sound source signal is colored, there is a disadvantage that the input audio signal and the output audio signal have different spectral structures and are unnatural.

SUMMARY OF THE INVENTION An object of the present invention is to provide a linear predictive speech analysis / synthesis apparatus which uses a sound source signal in which energy is not concentrated, and has a high sound quality because the input speech signal and the output speech signal have the same spectral structure.

[Means for solving the problem]

The linear predictive speech analysis / synthesis apparatus according to the present invention synthesizes a synthesizing means in which a cascade frequency characteristic of a spectrum envelope characteristic of the sound source information and a spectrum envelope frequency characteristic of the speech synthesis filter matches the spectrum envelope frequency characteristic of the input audio signal. As an example of specific means, information on the distinction between voiced and unvoiced sounds of the input voice signal at each time interval predetermined on the analysis side, information on the fundamental frequency when the input voice signal is a voiced sound The sound source information including the power information and the linear prediction coefficient indicating the spectral envelope or a coefficient equivalent to the linear side coefficient are measured, and the sound source information and the linear prediction coefficient or the coefficient equivalent to the linear prediction coefficient are calculated on the synthesis side. A linear predictive speech analysis / synthesis apparatus that synthesizes an output speech signal based on a pulse from a pulse generator and the line A digital filter to which a prediction coefficient or a coefficient equivalent to the linear prediction coefficient is inputted, and an excitation signal generating means for outputting a pulse train or a noise signal having a cycle equal to the cycle of the fundamental frequency based on the voiced / unvoiced discrimination information A transversal filter to which the output of the digital filter and the output of the excitation signal generating means are respectively input and have an impulse response obtained by inverting the impulse response of the digital filter with respect to time; and the transversal filter based on the power information. Sound source signal generating means for controlling the output of the sound source and outputting a sound source signal;
A lossy synthesis filter having a characteristic obtained by adding a predetermined loss to a filter characteristic to which the sound source signal is input and determined by the linear prediction coefficient or an equivalent coefficient of the linear prediction coefficient, wherein the linear prediction coefficient or the linear The quotient obtained by dividing the transfer function of the synthesis filter determined by a coefficient equivalent to the prediction coefficient by the completion of the transfer of the lossy synthesis filter is defined as the transfer function of the digital filter.

〔Example〕

 Next, the present invention will be described with reference to the drawings.

 FIG. 1 is a block diagram showing one embodiment of the present invention.

In the embodiment shown in FIG. 1, window cutout processors 1 and 2 for inputting an input audio signal, and an output signal of the window cutout processor 1 are input and K parameters k _{1 to} k _p and a power parameter pw are set. Output
LPC analyzer 3, K quantizer 4 for inputting K parameters k _{1 to} k _p, and power quantizer 5 for inputting power parameter pw
A pitch extractor 6 that receives an output signal of the window cutout processor 2 and outputs a pitch parameter pt;
, A K quantizer 4, a power quantizer 5, and a multiplexing circuit 8 for inputting output signals of the pitch quantizer 7, and a multiplexing circuit 8 via a transmission line (not shown). A separation circuit 9 for inputting an output signal of the conversion circuit 8, a K decoder 10, a power decoder 11, a pitch decoder 12, and a K decoder 10 each having an input terminal connected to the separation circuit 9. K / α converter 13 for inputting output K parameters k _{1 to} k _p and outputting α parameters α _{1 to} α _p
A sound source signal generation unit 14 that inputs the power parameter pw output from the power decoder 11, the pitch parameter pt output from the pitch decoder 12, and the α parameters α _{1 to} α _p , The output signal and the α parameters α _{1 to} α _p are provided with a lossy synthesis filter 15 for outputting an input audio signal. The part up to the multiplexing circuit 8 is the analysis side, and the part after the separation circuit 9 is the synthesis side.

The window cutout processor 1 cuts the output audio signal to a cutoff frequency of 3.4 kHz.
The bandwidth is limited by a low-pass filter, and then sampled at a sampling frequency of 8 kHz, each sample is quantized to a predetermined number of bits, and a window cutout that multiplies a 30 ms quantized sample by a Hamming window function is an analysis frame period. Is performed every 20 ms. The window cutout processor 2 processes the input audio signal in the same manner as the window cutout processor 1 except that a rectangular window function is used as the window function.

The LPC analyzer 3 performs linear prediction analysis (Iinear predictive) on the signal input from the window cutout processor 1 for each analysis frame.
coding: LPC) to extract K parameters k _{1 to} k _p , which are p-order partial autocorrelation coefficients, and a power parameter pw representing the power (1/2 power) of each analysis frame of the input speech signal. Is output. The pitch extractor 6 determines whether the signal input from the window cutout processor 2 is a voiced sound or an unvoiced sound for each analysis frame, and if the signal is a voiced sound, extracts a pitch period,
Judgment result and pitch cycle collectively pitch parameter pt
Output as

The K parameters k _{1 to} k _p , the power parameter pw, and the pitch parameter pt obtained by the LPC analyzer 3 and the pitch extractor 6 are K
After being quantized to a predetermined number of bits by a quantizer 4, a power quantizer 5, and a pitch quantizer 7, respectively, multiplexed by a multiplexing circuit 8, transmitted by a transmission line, and separated by a separating circuit 9. K
Decoder 10, the power Decoder 11 is decoding respectively pitch Decoder 12, again becomes K parameters k ₁ to k _p, power parameter pw, pitch parameter pt.

K / alpha converter 13, the alpha parameter alpha ₁ to? _P is the linear prediction coefficients calculated by known methods from the K parameter k ₁ to k _p, which is decoding, and outputs.

As described in detail above, the sound source signal generation unit 14 generates a sound source signal based on the reproduced α parameters α _{1 to} α _p , the pitch parameter pt, and the power parameter pw, and the lossy synthesis filter 15 Input and α parameters α _{1 to} α _p
The output audio signal is synthesized based on

 First, the lossy synthesis filter 15 will be described.

 FIG. 2 is a block diagram of the lossy synthesis filter 15.

The lossy synthesis filter 15 includes a subtractor 31 and p multipliers 32 for inputting a constant γ satisfying 0 <γ <1 from one input terminal.
And p number of delay circuits 33 each providing a delay equal to the sampling period in the window cutout processors 1 and 2, and p number of inputting α parameters α _i (i is an integer of 1 to p) from one input terminal , And an adder 35. The output terminal of one multiplier 32 is connected to the input terminal of one delay circuit 33, and each one of the multipliers thus connected is connected.
32 and p sets of cascade connection circuits each including a delay circuit 33 are cascade-connected to the output terminal of the subtractor 31. The output terminal of the i-th delay circuit 33 from the top is also connected to the other input terminal of the multiplier 34 for inputting the α parameter α _i . All the adders 35 add the multiplication calculation power of the multiplier 34. The subtractor 31 subtracts the added output of the adder 35 from the input sound source signal. Subtractor 31
Is branched and taken out as an output audio signal.

A circuit obtained by removing all the multipliers 32 from the lossy synthesis filter 15 with the constant γ set to 1, in other words, is a known LPC synthesis filter. The lossy synthesis filter 15 has a configuration in which a loss determined by a constant γ is given to each stage of the LPC synthesis filter, and its waveform response is a waveform response obtained by damping the waveform response of the LPC synthesis filter.

The transfer function H ₁ (Z) of the lossy synthesis filter 15 is Is represented by Usually, the transfer function H (Z) of the LPC synthesis filter used in the linear prediction type speech analysis / synthesis apparatus is Is represented by FIG. 4 shows an example of frequency transmission characteristics (spectral envelope characteristics) of H (Z) and H ₁ (Z), and FIG. 5 shows an example of impulse response. In addition, in FIG. 4 and FIG. 5, H ₁ (Z)
Shows the case where γ = 0.8. When this coefficient γ is 1.0, H ₁ (Z) = H (Z), and when γ = 0.0, H
The frequency transmission characteristic of ₁ (Z) is completely flat, and the impulse response is a unit pulse.

It should be noted that the multiplier 32 is entirely removed from the lossy synthesis filter 15 and the multiplier 34 has a value α _i γ ⁱ instead of the α parameter α _1.
Is input, a lossy synthesis filter having the same transfer function as the lossy synthesis filter 15 can be configured.

 Next, the sound source signal generation unit 14 will be described.

 FIG. 3 is a block diagram of the sound source signal generation unit 14.

The sound source signal generation unit 14 includes a clock generator 20, a pulse generator 21, a standard digital filter 22 that receives output signals of the clock generator 20, the pulse generator 21, and α parameters α _{1 to} α _p . A plurality of (the number of which will be described later) delay circuits 23 cascaded to the output terminal of the standard digital filter 22 and each receiving an output signal of the clock generator 20, and a pulse train generator 24 receiving a pitch parameter pt , A noise generator 25, a switch 26 for selectively outputting either the pulse train generator 24 or the noise generator 25 under the control of the pitch parameter pt, and sampling in the window cutout processors 1 and 2. Each of the delay circuits is provided with a delay equal to the cycle and connected in cascade to the output terminal of the switch 26, and the number of the delay circuits 27 is one less than the number of the delay circuits 23, and the delay circuits 23 and 23 arranged in the same order from the rear end. Yo A multiplier 28 having two inputs each of the output signal of the delay circuit 27 and a multiplier 28 having two inputs of the output signal of the delay circuit 23 arranged at the head and the input signal of the delay circuit 27 arranged at the beginning, An adder 29 that adds the multiplication power of the multiplier 28 and a multiplier 30 that inputs the power parameter pw and the added output of the adder 29 and outputs the multiplication power as a sound source signal are provided.

The pulse train generator 24 generates a pulse train with repetitive synchronization equal to the pitch period in the pitch parameter pt. The noise generator 25 outputs white noise such as an M sequence. Switch 26
Selects the output pulse train of the pulse generator 24 for voiced speech and the output noise of the noise generator 25 for unvoiced speech in accordance with the determination result in the pitch parameter pt, and outputs it as an excitation signal.

The part of the sound source signal generator 14 other than the pulse train generator 24, the noise generator 25, and the switch 26 is excited by the excitation signal output from the switch 26, and generates a sound source signal as described below.

(Α parameter α ₁ to LPC synthesis filter)
Transfer function H (Z) determined by α _p and transfer function (determined by α parameters α _{1 to} α _p ) of lossy synthesis filter 15
For H ₁ (Z), the transfer function is The standard digital filter 22 is configured so that
The clock generator 20 outputs as many clock pulses as the required impulse response length of the standard digital filter 22 for each analysis frame. The repetition period of this clock pulse is set sufficiently shorter than the sampling period in the window cutout processors 1 and 2. The pulse generator 21 outputs one impulse for each analysis frame. Each delay circuit
The delay circuit 23 is composed of, for example, a D-type flip-flop that uses a clock pulse output from the clock generator 20 as a clock to operate in parallel for a required number of bits.
Is equal to the number of clock pulses generated by the clock generator 20.

The α parameters α _{1 to} α _p are input for each analysis frame, the transfer function H ₂ (Z) of the standard digital filter 22 is set, and then the impulse is input from the pulse generator 21 and the clock pulse from the clock generator 20 is input. When the standard type digital filter 22 is operated and the clock pulse is completely output, a signal representing the impulse response of the standard type digital filter 22 is obtained at the output terminal of each delay circuit 23, and is held until the next analysis frame.

By the way, the transversal filter including the delay circuit 27, the multiplier 28, and the adder 29
Because of the fact that the impulse response of the standard digital filter 22 is temporally inverted, the impulse response has an impulse response based on the fact that the impulse response is obtained from 23 and the connection correspondence between each delay circuit 23 and each multiplier 28. The excitation signal output from the switch 26 is input to the transversal filter, and the output signal is correlated with the power of the input audio signal by the multiplier 30 and output to the lossy synthesis filter 15 as a sound source signal. Note that the multiplier 30 is connected to the switch 26
Can be connected immediately after

Now, the spectrum structure of the sound source signal output from the sound source signal generation unit 14 is represented by the transfer function H with the excitation signal output from the switch 26.
₂ Equivalent to the spectral structure of the output signal when the (Z) digital filter is excited. This sound source signal is the transfer function
Since the output audio signal passes through the lossy synthesizing filter 15 of H ₁ (Z), the spectrum structure of the output audio signal is such that the excitation signal has a transfer function H (Z) (= H ₁ (Z) · H ₂ (Z) ) Matches the spectral structure of the output signal that has passed through the LPC synthesis filter, and therefore matches the spectral structure of the input speech signal. In addition, the transfer function of the impulse response of the transversal filter that creates the sound source signal from the excitation signal is H
_{By making} the impulse response of the digital filter which is ₂ (Z) into an inverted impulse response, the phase relationship of the process of generating the output audio signal from the excitation signal becomes the phase of the process of the transfer function H (Z) processed by the LPC synthesis filter. Unlike the relationship, even when the excitation signal is a pulse train, the energy of the output audio signal does not concentrate on the pitch excitation point.

〔The invention's effect〕

As described above, the present invention provides a lossy synthesis filter and an impulse of a digital filter for synthesizing an output audio signal that matches a spectrum structure of an input audio signal from an excitation signal that is a pulse train or a noise signal together with the lossy synthesis filter. By providing a sound source signal generating means for generating a sound source signal from an excitation signal by a sound source waveform generating means having an impulse response whose response is inverted, energy is not concentrated at a pitch excitation point, and furthermore, an input sound signal and an output sound signal Therefore, there is an effect that a linear prediction type speech analysis / synthesis apparatus having a good sound quality by matching the spectrum structures can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a block diagram of a lossy synthesis filter 15 in the embodiment shown in FIG. 1, FIG. FIG. 4 is a waveform diagram for explaining the spectral envelope characteristics of the lossy synthesis filter 15, and FIG. 5 is a waveform diagram for explaining the impulse response characteristics of the lossy synthesis filter 15. 1,2 ... window cutout processor, 3 ... LPC analyzer, 6 ... pitch extractor, 13 ... K / α converter, 14 ... sound source signal generator, 15
…… Synthesis filter with loss, 20 …… Clock generator, 21…
… Pulse generator, 22 …… Standard digital filter, 2
3,27 delay circuit, 24 pulse train generator, 25 noise generator, 26 switch, 28, 30 multiplier, 29 adder, 31 subtractor, 32, 34: Multiplier, 33: Delay circuit, 35: Adder.

Claims (2)

(57) [Claims] An input speech signal is analyzed to extract voiced / unvoiced discrimination information, sound source information including fundamental frequency information when voiced, and a linear prediction coefficient indicating a spectrum envelope or a coefficient equivalent thereto. A linear predictive speech analysis / synthesizer for transmitting the signal to the synthesizing side and synthesizing the audio signal on the synthesizing side, wherein the synthesizing side receives an excitation signal generated based on the sound source information, and Sound source waveform shaping means for forming a sound source signal by involving an impulse response determined by a prediction coefficient or a coefficient equivalent thereto; A lossy synthesis filter having a characteristic obtained by adding a predetermined loss to the sound source signal, and a spectrum envelope frequency characteristic of the sound source wave signal and a loss of the lossy synthesis filter. Linear predictive speech analysis and synthesis device, characterized in that fitted to the spectrum envelope frequency characteristics of the input audio signal cascade frequency characteristics of the spectrum envelope frequency characteristics. 2. The sound source information including the voiced / unvoiced discrimination information of the input voice signal, the fundamental frequency information and the power information when the voice signal is voiced, and the spectrum envelope at each time interval predetermined on the analysis side. A linear prediction type speech which measures a linear prediction coefficient or a coefficient equivalent to the linear prediction coefficient, and synthesizes an output audio signal based on the sound source information and the linear prediction coefficient or a coefficient equivalent to the linear prediction coefficient on a synthesis side. In the analysis / synthesis apparatus, the synthesis side includes a digital filter to which the pulse from the pulse generator and the linear prediction coefficient or a coefficient equivalent to the linear prediction coefficient are input, and the voiced / unvoiced sound discrimination information. An excitation signal generating means for outputting a pulse train or a noise signal having a cycle equal to the cycle of the fundamental frequency; an output of the digital filter and the excitation signal A transversal filter to which the output of the generating means is input and which has an impulse response obtained by temporally inverting the impulse response of the digital filter, and a sound source for controlling the output of the transversal filter based on the power information and outputting a sound source signal Signal generating means; and a lossy synthesis filter having a characteristic obtained by adding a predetermined loss to a filter characteristic to which the sound source signal is input and determined by the linear prediction coefficient or a coefficient equivalent to the linear prediction coefficient. A quotient obtained by dividing the transfer function of the synthesis filter determined by the prediction coefficient or a coefficient equivalent to the linear prediction coefficient by the transfer function of the lossy synthesis filter is used as the transfer function of the digital filter, and an output audio signal is obtained from the synthesis filter. A linear predictive speech analysis / synthesis apparatus characterized by the following.