JPH041920B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a pitch extraction device that performs pitch extraction based on an autocorrelation coefficient obtained by analyzing a speech waveform with a frame period approximately equal to the pitch period. This relates to a pitch extraction device that can be reduced in size. It is known that the voiced part of a speech waveform has a periodic repeating waveform, and the change characteristic of the period (pitch period) is an important parameter in speech analysis, synthesis, recognition, etc. For example, in a speech analysis and synthesis system, the pitch extraction result extracted by the analysis section has a large effect on the quality of synthesized speech synthesized by the synthesis section. Conventionally, the pitch period extraction method for audio waveforms is
Methods using various analysis parameters are known, such as a method of calculating and extracting an autocorrelation coefficient for each frame having a time length of approximately the pitch period. Pitch extraction methods based on autocorrelation coefficients are widely used because the autocorrelation coefficients can be obtained through processing in the time domain, and the influence of the phase between the analyzed waveform and the frame is relatively small. . However, pitch extraction methods based on autocorrelation coefficients often erroneously detect periods that are an integral multiple of the pitch period or N1/N2 times the pitch period as the pitch period (however, N1 and N2 are integers, N1 < N2).
The above-mentioned problem in the pitch extraction method where a cycle that is an integer multiple of the original pitch cycle is mistakenly detected as a pitch cycle (hereinafter referred to as "integer multiple pitch cycle error") is solved by the method proposed by the present inventor in Japanese Patent Application Laid-Open No. 54−139307
It is described in detail in ``Pituchi Extraction Device.'' Many pitch extraction methods based on autocorrelation coefficients limit the pitch search range to around the pitch period in order to alleviate integer multiple pitch period errors. Conventionally, the pitch search range determination method divides speech into voiced sections and unvoiced sections, and then sets the pitch search range for the voiced section after unvoiced sections have been connected for a relatively long time to an initial setting value, or The pitch search range for a voice section has been set in the vicinity of the past pitch period. This type of conventional pitch search range determination method is described in detail in the present application, ``Pitch Extraction Apparatus,'' proposed by the inventor of the present invention, ``Pitch Extraction Apparatus'' published in Japanese Patent Application Laid-Open No. 56-42296. However, in the conventional method, an initially set pitch search range is given for a voiced section after an unvoiced section of a predetermined period or more, but the previous pitch period is held for an unvoiced section of a shorter period. It was processed in the same way as voiced intervals. In fact, even if there is a short-term unvoiced sound section, the pitch period often changes before and after it, and in conventional two-way processing, pitch period extraction after a short-term unvoiced sound section is not accurate. It had become something lacking. SUMMARY OF THE INVENTION An object of the present invention is to provide a pitch extraction device capable of accurate pitch retrieval even when short-time unvoiced sections are mixed in generally voiced sections. The present invention has means for controlling and expanding the pitch search range in accordance with the length of a short unvoiced sound section mixed in a generally voiced sound section. Next, embodiments of the present invention will be described with reference to the drawings. The figure is a block diagram for explaining an embodiment of the present invention. An input audio signal is input to an A/D converter 2 via an audio input terminal 1 . A/D converter 2
samples the input audio signal at, for example, 8kHz and quantizes each sample. The sampled and quantized audio signal is written into the buffer memory 3. The buffer memory 3 stores data once every 20 msec after the frame period, for example.
Speech samples corresponding to the length of the analysis window, for example 30 msec, are output to the voiced/unvoiced discriminator 4 and the autocorrelation coefficient measuring device 5. The voiced/unvoiced discriminator 4 determines whether the voice is voiced or unvoiced.
It is determined whether there is no voice, and the determination result is further output to the pitch search device 6. The autocorrelation coefficient measuring device 5 calculates the autocorrelation sequence of the audio waveform supplied from the buffer memory 3 for each frame period into a pitch period distribution range, e.g.
Measure the delay range corresponding to 2.6msec to 15msec. The autocorrelation coefficient ρ _P of the delay τ measured by the autocorrelation coefficient measuring device 5 is It is. However, x(i) is the i-th audio sample in the analysis frame, N is a natural number, for example
Number equivalent to 15msec (120 for 8kHz sample)
It is. The measured autocorrelation coefficient sequence is supplied to a pitch searcher 6. The pitch searcher 6 searches for the maximum value in the pitch search range of the autocorrelation coefficient sequence supplied from the autocorrelation coefficient measuring device 5, and determines the delay time τ _P corresponding to the maximum value. Information representing the pitch search range is supplied from a pitch search range measuring device 10. The pitch searcher 6 further _{sets the above τ P to "0" if the voiced/unvoiced discrimination result supplied from the voiced/unvoiced discriminator 4 is voiced, and sets the above τ P to} "0" if the discrimination result is unvoiced.
Change it to and output it. The output of the pitch searcher 6 is supplied to a continuous voicing detector 7, a continuous unvoiced detector 8, a pitch search range measuring device 10, and a pitch output terminal 11. The continuous voicing detector 7 receives data supplied from the pitch searcher 6 for a relatively long time, for example, 100 msec.
If it is not "0" continuously, that is, if a continuous portion of voiced sound is detected, a trigger is output and the flip-flop 9 is set to "1". The continuous unvoiced detector 8 outputs a trigger when the data supplied from the pitch searcher 6 is "0" continuously for a relatively long time, for example, 300 msec, that is, when a continuous portion of unvoiced sounds is detected,
Reset flip-flop 9 to "0". The pitch search range measuring device 10 is a storage program type processing device, and when the signal supplied from the flip-flop 9 is "0", the pitch search range measuring device 10 sets an initial pitch search range, for example, 2.625 msec to 11.25 msec. Output the setting value to. Also, when the signal supplied from the flip-flop 9 is "1", the past pitch period supplied from the pitch searcher 6 is leaked integrated to calculate the average pitch period, and then the average pitch period is calculated. The pitch search range is set and the set value is output to the pitch search device 6. When the signal supplied from the flip-flop 9 changes from "0" to "1", that is, in the transition from unvoiced to voiced continuation, the pitch search range measuring device 10 detects the past few frames supplied from the pitch search device 6. The average value of the pitch periods of minutes, for example, the average value of the pitch periods of 100 msec, is calculated, a pitch search range is set in the vicinity of the average pitch period, and the set value is output to the pitch search device 6. When the output of the flip-flop 9 is "1", that is, in the interval set as a continuous interval of voiced sounds, voiced/
When the unvoiced discriminator 4 performs unvoiced determination for several frames and then performs voiced determination, the pitch search range measuring device 10 performs a characteristic process. When the voiced/unvoiced discriminator 4 performs voiceless discrimination when the output of the flip-flop 9 is "1", the output of the pitch search device 6 becomes "0", so the pitch search range measuring device 10
does not perform a simple leaked integration of the pitch cycle. The pitch search range measuring device 10 has pitch cycle information supplied from the pitch search device 6 as “0”.
The leaked integrated average pitch period just before is held. Further, the pitch search range measuring device 10 widens the pitch search range from the held average pitch cycle to correspond to the interval where the pitch cycle is "0", that is, the silent interval. Assuming that the average pitch period immediately before becoming an unvoiced sound is T _A , the coefficient that determines the upper limit of the pitch search range is A (A > 1.0), and the coefficient that determines the lower limit of the pitch search range is B (B < 1.0),

【table】

Claims (1)

[Claims] 1. In a pitch extraction device that discriminates voiced/unvoiced parts of an input audio signal and extracts a pitch period from an autocorrelation sequence of the voiced sound section, means for determining the length of a short unvoiced sound section mixed in the voiced sound section; A pitch extraction device characterized by comprising means for expanding and controlling a pitch search range according to the length of the unvoiced sound section.