JPS5912191B2 - Speech synthesis method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech synthesis method for synthesizing human speech using an extremely simple structure, and in particular, a method for collecting various data (coefficients, etc.) by performing a predetermined analysis of the original speech. I! The present invention relates to a speech synthesis method in which a digital signal is determined by an information processing device, is logarithmically compressed, and then analog-to-digital converted into a digital signal, and when speech is synthesized from such a digital signal, logarithmic expansion is performed to widen the dynamic range.

In general, speech synthesis involves converting original speech into an electrical signal using a microphone 1, as shown in FIG. The recorded low-frequency signal is applied to a low-pass filter 4 via the low-frequency amplifier 2, logarithmically compressed by a logarithmic compression circuit 5, and then converted to an analog-to-digital (hereinafter abbreviated as A/D) converter 6. digitize (PCM).

The converted digital signal is sampled from the audio signal VA in Figure 3 at intervals (for example, every 20 ms) and analyzed, and then various data such as PARCOR analysis, which is currently widely used, is analyzed. Then, data regarding voiced and unvoiced sounds, amplitude data, parameters, etc. are determined by a parameter storage device L provided as a centralized information processing means, and stored in a memory element (drum, ROM, etc.) 8.

Therefore, since various data extracted at each predetermined interval are required, the memory capacity used per second is determined by the amount of bits, and generally, it is 2400bps.
) If 4800 bps or 9600 bps is used, and 2400 bps is taken as an example, 2400 bits of memory (to store various data) is required to generate one second of audio.

The problem here is bit allocation to parameters etc. as the amplitude data filter coefficient data.
Based on various experiments and theories up to now, it has been determined to some extent how to allocate the memory in order to use the memory most effectively, and according to this, in the case of the above-mentioned 240 Obps, the allocation to the amplitude data is 6 bits. be.

Therefore, it is difficult to quantize (digitize) human voice amplitude with 6 bits due to dynamic range, so as shown in Figure 1, the original voice is compressed to some extent and used to extract amplitude data. There is a need.

This pattern is an example in which the compression ratio is set to 1/2 (DB), and is designated as P in the input/output characteristics of FIG. In FIG. 4, L indicates a linear characteristic, and E indicates an elongated characteristic. When synthesizing speech, even if it is synthesized using the same amplitude data, the dynamic range will be audibly felt to be narrowed to some extent.

In this case, since each data (parameter) is extracted after 1 compression, the noise level also increases.

2. The synthesized sound lacks some dynamic range, resulting in an unnatural sound.

Considering these factors, it is absolutely necessary to get the sound close to the original sound.

For this reason, the present invention provides a novel speech synthesis method that eliminates the above-mentioned drawbacks by further providing a logarithmic expansion circuit when original speech is digitized using the configuration shown in FIG. It provides
This will be explained below with reference to the drawings.

Figure 2 is a block diagram for explaining the speech synthesis method of the present invention, Figure 5 A, the mouth is a waveform diagram of each part in Figure 2, and Figure 6 is a block diagram for explaining the speech synthesis method of the present invention.
The figure shows an explanatory waveform diagram when obtaining various data in voice analysis.

In FIG. 2, 9 is a noise generator, 10 is a pulse generator,
11 is a switching means for switching the output of the noise generator or pulse generator to derive a predetermined signal; 12 is an amplification stage; 13
is a digital filter, and 14 is a digital analog (D
/A) converter, 15 is a logarithmic expansion circuit, 16 is a low frequency amplifier, 17 is a sound emitting means such as a speaker, 7 is a parameter storage device provided as a centralized information processing means, and the switching means 1
1 amplification stage 12 and digital filter 13, respectively.
It is configured to supply data signals consisting of silent switching data, amplitude data, and filter coefficients. When synthesizing human speech, the data (parameters) representing the characteristics of the speech include information on whether the sound is voiced or unvoiced, pitch frequency and amplitude, and information on the transmission of the human vocal tract. PARCOR as data representing characteristics
The method employs the k parameter.

Next, we consider the speech waveform as the output of the vocal tract excited by the sound source, and assume that the characteristics of the vocal tract are passive and linear.
A linear predictive analysis is performed on the speech waveform shown in the figure, and if the input of the sound source is ε(n) and the output is X(n), then the transfer function H(z) from the input ε(n) to the output X(n) is When transformed into z,
It is expressed by the following formula.

In the above formula, σ represents a constant.

When the sampling interval t is determined from this, a value of 20 ms is considered to be appropriate as described above.

Next, when a human voice signal is logarithmically compressed by the logarithmic compression circuit 5 in the configuration shown in FIG.
Set it to 5% or 3.4KHz. This is equal to the band of the filter 4 which functions to prevent aliasing noise in FIG. This value is adopted because the optimum sampling frequency is 8 KHz, which is twice the above-mentioned 4 KHz, that is, the period is 125 μs. In FIG. 2, the first
A white nozzle from the noise generator 9 shown in FIG.
By appropriately switching the pulses from the pulse generator 10 shown in FIG. A low frequency signal is obtained through the converter 14 and is logarithmically compressed during speech analysis.
Since the signal is logarithmically expanded and returned to its original state in the logarithmic expansion circuit 15 as shown by characteristic E in FIG. 4, it is possible to suppress the increase in the noise level described above and also prevent a lack of dynamic range.

As described above, according to the present invention, when performing speech synthesis using various data obtained through logarithmic compression during speech analysis, the original speech can be restored by logarithmic expansion and approximated by the original speech, According to the method of the present invention, an extremely natural synthesized sound can be obtained without increasing the memory capacity, and the method of the present invention has a wide range of applications for various controls, alarms, etc.

[Brief explanation of the drawing]

1 and 2 are block diagrams for explaining the speech synthesis method of the present invention, FIG. 3 is a waveform diagram of the original speech, and FIG. 4 is a logarithmic compression and logarithm diagram for explaining the method of the present invention. FIG. 5 shows an input/output characteristic diagram representing the expansion characteristic, and FIG. 5 shows a waveform diagram of each part in FIG. 1. FIG. Explanation of main figure numbers, 5... Logarithmic compression circuit, 6...
...A/D converter, I...Parameter storage device, 8...Memory element, 9...Noise generator, 10...Pulse Generator, 11...
...Switching means, 12...Amplification stage, 13...
... Digital filter, 14 ... D/A converter, 15 ... Logarithmic expansion circuit, 16 ... Low frequency amplifier, 17 ... Sound emitting means.

Claims (1)

[Claims] 1. A switching means connected to the output side of the noise generator and the pulse generator, an amplification stage, a digital filter, a digital-to-analog converter, a logarithmic expansion circuit, a low frequency amplifier, and It consists of a sound emitting means, and a centralized information processing means connected to the switching means, an amplification stage, and a digital filter. The switching data, amplitude data, and filter coefficient data of each stage and digital filter are determined and logarithmically compressed, each of the data is applied to the means, the amplification stage, and the digital filter, respectively, and the logarithmic expansion circuit converts the data into digital data. Logarithmically expands the output of the analog conversion circuit,
A voice synthesis method characterized by generating a low frequency signal from a sound emitting means.