RU2174714C2 - Method for separating the basic tone - Google Patents

Abstract

FIELD: automation engineering. SUBSTANCE: method involves filtering-out the basic maximum points of time-dependent function of initial speech signal. Maximum search is to be blocked for single blocking interval after the selected maximum. Descending signal is formed after the blocking pause is over lowering relative to the selected maximum. The descending signal is generated with its maximum value exceeding the absolute maximum of the basic signal over the basic tone time interval under analysis. EFFECT: higher reliability in detecting basic maximums of the initial signal time function. 7 cl, 154 dwg, 15 tbl

Description

 The present invention relates to the field of primary signal extraction of a signal, in particular to the field of extraction of the fundamental tone of a speech signal.

 Known methods for highlighting the fundamental tone by the time or spectral function of the original signal.

 In the filter method for isolating the fundamental tone [M. A. Sapozhkov. Speech signal in cybernetics and communication. - M .: Svyazizdat, 1963. - 472 p.] The original signal is limited by a low-pass filter or a band-pass filter so that the signal after filtering contains mainly only the main harmonic of the signal. From the filtered signal, the main maxima corresponding to the moments of excitation of the vocal tract are determined. The method does not track fast changes in the frequency of the fundamental tone, for example, at transitions between voiced and unvoiced sounds, it requires precise setting of the cutoff frequency of the low-pass preliminary filter or the band-pass filter.

 In the Rabiner – Gold method [V. Gold, L. Rabiner. Parallel processing techniques for estimating pitch period of speech in the time domain // J. Acoustic Soc. Am., 1969 - 46.- N 2 (Pt. 2). - P. 442 - 448] the speech signal is limited by a low-pass filter with a cutoff frequency in the range of 600 - 800 Hz, the distances between all signal maxima, the distances between all signal minima, the distance between the differences between the minima and maxima are determined in the filtered signal, and then the period is determined by the greatest number of matches of the selected distances. The method does not determine the moments of excitation of the speech tract, has an increased number of errors with rapid changes in the speech signal.

 In the cepstral method for isolating the fundamental tone [A.M. Noll. Short-Time spectrum and "Cepstrum" techniques for vocal-pitch detection // J. Acoustic. Soc. Am. - 36. - N 2, 1955] receive the broadband spectrum of the signal by the direct Fourier transform, logarithm the spectrum, perform the inverse Fourier transform of the logarithmic peak, obtained after the inverse Fourier transform of the signal located in the region of valid values of the period of the fundamental tone, determine the period of the fundamental tone . The method requires a large number of calculations, it does not determine the moments of excitation of the speech path, has a large number of errors during sudden changes in the input signal, including the transitions between voiced and unvoiced sounds, and is sensitive to broadband interference.

 In the autocorrelation method for isolating the fundamental tone [A. A. Pirogov. Device for automatically highlighting the fundamental tone. - A. p. N 129 739 USSR. - Prior, from 08.06.58. - NKI 21e 1/20 42e] determine the autocorrelation function of the input signal, at the maximums of which they decide on the period of the fundamental tone. The autocorrelation method is sensitive to changes in the shape of the envelope of the temporal function of the original signal, to changes in the waveform from period to period of the fundamental tone, has failures in harmonics and subharmonics of the fundamental tone, does not determine the moments of excitation of the speech path, requires a large number of calculations.

 In the LLK algorithm (Lobanova - Levin - Koval) [Methodical recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. S. Petersburg. - St. Petersburg, 1997. - S. 279] first determine the signal autocorrelation, find the most probable value of the period of the fundamental tone, then use the peak method to determine the signal maximums taking into account the values of the most probable period of the fundamental tone. This method gives an increased number of errors with sudden changes in the fundamental tone, especially in the areas of transitions between voiced and unvoiced sounds.

 Of the known methods, the closest in technical essence is the peak method of isolating the fundamental tone [L. O. Dolansky. Instantaneous pitch period indicator // J. Acoust. Sos. Am. - 27. - N 11, 1955. - P. 67 - 72], in which the main maxima of the initial signal are determined by sequentially searching for the main maximum, after determining the main maximum, the blocking interval is set during which the maximum is not searched for at the end of the blocking time they generate a signal falling relative to the last found maximum, at the moment the original signal exceeds the falling signal, the search begins for the next main maximum of the original signal, after which the process is repeated.

The specified peak method has several significant disadvantages. It gives big errors (failures) at the moment of appearance of false maxima after the blocking time, comparable in magnitude with the main maxima. The appearance of false maxima is due to the following main reasons:
1. A speech signal is a non-stationary process. If in stationary sections its structure is more or less regular, then in transition sections between phonemes, between pauses and speech, it changes both in form and in intensity. The number of false maxima in non-stationary areas as compared with stationary increases [Vocoder Telephony / Ed. A.A. Pirogov. - M.: Communication, 1974. - 536 pp.] Due to changes in the source spectrum, due to changes in the shape of the vocal tract and due to non-linear effects in the source and vocal tract.

2. The transmission characteristic of the vocal tract can emphasize the higher harmonics of the fundamental tone. So the phoneme / and / has a maximum transfer characteristic in the region of the first formant F ₁ = 240 Hz. If the frequency of the fundamental tone lies in the range of 100 - 120 Hz, then the second harmonic of the fundamental tone falls into the region of the first formant and is emphasized. The emphasized second harmonic causes malfunctions at twice the frequency of the fundamental tone.

 3. In the process of speech formation, additional excitations of the vocal tract may occur within the period of the fundamental tone [J.N. Holmes. An investigation of the volume velocity waveform at the larinx during speech by means of inverse filter // Proc. Speech Cmmun. Seminar - Stockholm, 1962. - Vol 1. - B4].

 4. False maxima may appear due to phase shifts between the frequencies of the formants [V. B. Gitlin et al. Choice of the interval for measuring the frequency and width of formants // Automatic recognition of auditory images: Abstract. doc. and message 10 All-Union. seminar schools. - Tbilisi: Metsniereba, 1978. - S. 20 - 21].

 5. False maxima may appear due to external additive interference.

 The presence of false maxima leads to the fact that the reliability of the allocation of the fundamental tone by the peak method depends on the shape of the speech-forming tract. With changes in the shape of the original signal, in particular, at transitions between phonemes, at transitions between voiced and unvoiced sounds, at transitions from pause to speech, the number of erroneous decisions increases. The reliability of the allocation of the fundamental tone by the peak method is sensitive to additive noise. The method requires a preliminary assessment of the presence of vocalization in the original signal. The reliability of the method is reduced in the presence of nonlinear constraints such as limit constraints (clipping).

 The aim of the invention is to increase the reliability of the allocation of the fundamental tone of the source signal by more reliable selection of the main maxima corresponding to the main excitation, when the shape of the original signal undergoes significant distortion of the form, including due to the appearance of false maxima within the period of the fundamental tone.

 This goal is achieved by the fact that in the known method, in which the main maxima of the time function are isolated, after the selected maximum during the blocking interval, no new maximum is searched, and after the blocked interval, the search for a new maximum is performed using a signal decreasing relative to the selected maximum to search of each subsequent maximum, a falling signal is generated, the maximum value of which at the end of the blocking time exceeds the absolute maximum of the initial signal ala at corresponding pitch period.

 In addition, in order to simplify the generation of the falling signal, the initial signal is preliminary, before searching for the maxima, logarithm, and the falling signal is constructed in the form of a linearly decreasing signal. The blocking time and the rate of decline of the falling signal are adapted to the current period of the fundamental tone. The source signal before the logarithm is smoothed using a low-pass filter or a band-pass filter.

 In order to ensure the stability of the correlation linear prediction method, the speech analysis by the correlation method is performed synchronously with the fundamental tone, the analysis interval is defined as the interval between the main maxima of the original signal marked by the proposed method for extracting the fundamental tone, and the original signal is multiplied by the decaying exponent between the selected maxima.

 When deciding on the presence of vocalization, the analysis of the initial signal is carried out within the interval between the maxima of the initial signal selected by the proposed method.

The proposed method is illustrated by drawings, where
in FIG. 1 is a timing chart explaining the algorithm of the proposed method for finding the main maxima of the source signal;
in FIG. 2 presents a list of operations by the proposed method;
in FIG. Figure 3 shows the oscillogram of the sentence with the transcription "We didn’t see such a net", pronounced by the announcer male VBG;
in FIG. 4 shows the reference pitch circuit obtained by manual processing of the waveform shown in FIG. 3;
in FIG. 5 shows the outline of the pitch of the waveform shown in FIG. 3 obtained by the peak method with a standard set of parameters corresponding to the male voice;
in FIG. 6 shows the outline of the pitch of the waveform shown in FIG. 3 obtained by the filter method with a standard set of parameters corresponding to the male voice;
in FIG. 7 shows the outline of the pitch of the waveform shown in FIG. 3 obtained by the Rabiner-Gold method with a standard set of parameters corresponding to the male voice;
in FIG. 8 shows the outline of the pitch of the waveform shown in FIG. 3, obtained by the cepstral method with a standard set of parameters corresponding to the male voice;
in FIG. 9 shows the outline of the pitch of the waveform shown in FIG. 3 obtained by the autocorrelation method with a standard set of parameters corresponding to the male voice;
in FIG. 10 shows the outline of the pitch of the waveform shown in FIG. 3 obtained by the LLK method with a standard set of parameters corresponding to the male voice;
in FIG. 11 shows the outline of the pitch of the waveform shown in FIG. 3 obtained by the proposed method with a standard set of parameters corresponding to the male voice;
in FIG. Figure 12 shows the oscillogram of the sentence “We didn’t see such a net”, pronounced by the announcer VBG man with pause marks at the intervals of the bows of the explosive deaf phonemes / t / and / c / in the word “such”;
in FIG. 13 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the peak method after adjusting the parameters of the method;
in FIG. 14 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the filter method after adjusting the parameters of the method;
in FIG. 15 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the Rabiner-Gold method after adjusting the parameters of the method;
in FIG. 16 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the cepstral method after adjusting the parameters of the method;
in FIG. 17 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the LLK method after adjusting the parameters of the method;
in FIG. 18 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the autocorrelation method after adjusting the parameters of the method;
in FIG. 19 shows the outline of the pitch of the waveform shown in FIG. 12 obtained by the proposed method with manual adjustment of pauses;
in FIG. Figure 20 shows an oscillogram of the sentence with the transcription "We didn’t see such a net", delivered by the announcer of the IGF woman;
in FIG. 21 shows a reference pitch outline obtained by manually processing the waveform shown in FIG. 20;
in FIG. 22 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the peak method;
in FIG. 23 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the filter method;
in FIG. 24 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the Rabiner-Gold method;
in FIG. 25 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the cepstral method;
in FIG. 26 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the autocorrelation method;
in FIG. 27 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the LLK method;
in FIG. 28 shows the outline of the pitch of the waveform shown in FIG. 20 obtained by the proposed method;
in FIG. Figure 29 shows an oscillogram of the spoken word "three" isolated, announcer male VBG;
in FIG. 30 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the peak method;
in FIG. 31 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the filter method;
in FIG. 32 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the Rabiner-Gold method;
in FIG. 33 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the cepstral method;
in FIG. 34 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the autocorrelation method;
in FIG. 35 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the LLK method;
in FIG. 36 shows the outline of the pitch of the waveform shown in FIG. 29 obtained by the proposed method;
in FIG. 37 shows, in a time-enlarged scale, the oscillogram of the transition from the bow / t / to the bell / p / is the beginning of the vowel / and / for the word “three” shown in FIG. 29;
in FIG. 38 shows the outline of the pitch of the waveform shown in FIG. 37 obtained by the proposed method;
in FIG. 39 is a time-enlarged waveform of the stationary portion of the vowel / and / for the word “three” shown in FIG. 29;
in FIG. 40 shows the outline of the pitch of the waveform shown in FIG. 39 obtained by the proposed method;
in FIG. 41 is a time-enlarged waveform of the ending of the vowel / and / for the word “three” shown in FIG. 29;
in FIG. 42 shows the outline of the pitch of the waveform shown in FIG. 41 obtained by the proposed method;
in FIG. Figure 43 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such” at a signal / noise ratio of 40 dB;
in FIG. 44 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the peak method;
in FIG. 45 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the filter method;
in FIG. 46 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the Rabiner-Gold method;
in FIG. 47 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the cepstral method;
in FIG. 48 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the autocorrelation method;
in FIG. 49 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the LLK method;
in FIG. 50 shows the outline of the pitch of the waveform shown in FIG. 43 obtained by the proposed method;
in FIG. Figure 51 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such” at a signal / noise ratio of 30 dB;
in FIG. 52 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the peak method;
in FIG. 53 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the filter method;
in FIG. 54 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the Rabiner-Gold method;
in FIG. 55 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the cepstral method;
in FIG. 56 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the autocorrelation method;
in FIG. 57 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the LLK method;
in FIG. 58 shows the outline of the pitch of the waveform shown in FIG. 51 obtained by the proposed method;
in FIG. 59 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such” at a signal / noise ratio of 20 dB;
in FIG. 60 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the peak method;
in FIG. 61 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the filter method;
in FIG. 62 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the Rabiner-Gold method;
in FIG. 63 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the cepstral method;
in FIG. 64 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the autocorrelation method;
in FIG. 65 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the LLK method;
in FIG. 66 shows the outline of the pitch of the waveform shown in FIG. 59 obtained by the proposed method;
in FIG. 67 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a male speaker VBG with pause marks at the intervals of the bows of explosive deaf phonemes / t / and / c / in the word “such” at a signal / noise ratio of 10 dB;
in FIG. 68 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the peak method;
in FIG. 69 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the filter method;
in FIG. 70 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the Rabiner-Gold method;
in FIG. 71 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the cepstral method;
in FIG. 72 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the autocorrelation method;
in FIG. 73 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the LLK method;
in FIG. 74 shows the outline of the pitch of the waveform shown in FIG. 67 obtained by the proposed method;
in FIG. Figure 75 shows the oscillogram of the sentence “We didn’t see such a net”, pronounced by a male speaker VBG with pause marks at the intervals of the bows of explosive deaf phonemes / t / and / c / in the word “such” at a signal / noise ratio of 5 dB;
in FIG. 76 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the peak method;
in FIG. 77 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the filter method;
in FIG. 78 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the Rabiner-Gold method;
in FIG. 79 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the cepstral method;
in FIG. 80 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the autocorrelation method;
in FIG. 81 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the LLK method;
in FIG. 82 shows the outline of the pitch of the waveform shown in FIG. 75 obtained by the proposed method;
in FIG. Figure 83 shows the oscillogram of the sentence “We didn’t see such a net”, pronounced by VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such” at a signal-to-noise ratio of 0 dB;
in FIG. 84 shows the outline of the pitch of the waveform shown in FIG. 83 obtained by the peak method;
in FIG. 85 shows the outline of the pitch of the waveform shown in FIG. 83 obtained by the filter method;
in FIG. 86 shows the outline of the pitch of the waveform shown in FIG. 83 obtained by the Rabiner-Gold method;
in FIG. 87 shows the outline of the pitch of the waveform shown in FIG. 83 obtained by the proposed method;
in FIG. 88 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such”, filtered by a low-pass filter with a cutoff frequency of 800 Hz;
in FIG. 89 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such” at a signal-to-noise ratio of 0 dB;
in FIG. 90 shows the outline of the pitch of the waveform shown in FIG. 89 obtained by the proposed method;
in FIG. 91 is a time-enlarged waveform of the transition from the bow / d / to the vowel / a / for the word “not seen” shown in FIG. 88;
in FIG. 92 shows the outline of the pitch of the waveform shown in FIG. 91 obtained by the proposed method;
in FIG. 93 is a time-enlarged waveform of the transition from the vowel / a / to the sonoric / l / and then to the vowel / and / for the word "not seen" shown in FIG. 88;
in FIG. 94 shows the outline of the pitch of the waveform shown in FIG. 93 obtained by the proposed method;
in FIG. 95 is a time-enlarged waveform of the transition from the sonoric / m / to the vowel / s / for the word “we” shown in FIG. 88;
in FIG. 96 shows the outline of the pitch of the waveform shown in FIG. 95 obtained by the proposed method;
in FIG. 97 is a time-enlarged waveform of the transition from the bow / t / to the vowel / a / and further to the bow / k / for the word "such" shown in FIG. 88;
in FIG. 98 shows the outline of the pitch of the waveform shown in FIG. 97 obtained by the proposed method;
in FIG. 99 is a time-enlarged waveform of the transition from the vowel / o / to the bow / d / for the word “net” shown in FIG. 88;
in FIG. 100 shows the outline of the pitch of the waveform shown in FIG. 99 obtained by the proposed method;
in FIG. 101 shows the oscillogram of the sentence “We didn’t see such a net”, pronounced by the announcer VBG man with pause marks at the intervals of the bows of the explosive deaf phonemes / t / and / c / in the word “such”;
in FIG. 102 shows the outline of the pitch of the waveform shown in FIG. 101 obtained by the proposed method;
in FIG. 103 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t / and / c / in the word “such”, the signal is limited to 100 samples;
in FIG. 104 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the peak method;
in FIG. 105 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the filter method;
in FIG. 106 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the Rabiner-Gold method;
in FIG. 107 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the cepstral method;
in FIG. 108 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the autocorrelation method;
in FIG. 109 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the LLK method;
in FIG. 110 shows the outline of the pitch of the waveform shown in FIG. 103 obtained by the proposed method;
in FIG. 111 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the intervals of the bows of explosive deaf phonemes / t / and / c / in the word “such,” the signal is limited to 50 samples;
in FIG. 112 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the peak method;
in FIG. 113 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the filter method;
in FIG. 114 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the Rabiner-Gold method;
in FIG. 115 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the cepstral method;
in FIG. 116 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the autocorrelation method;
in FIG. 117 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the LLK method;
in FIG. 118 shows the outline of the pitch of the waveform shown in FIG. 111 obtained by the proposed method;
in FIG. 119 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the bow intervals of explosive deaf phonemes / t and / c / in the word “such,” the signal is limited to 10 samples;
in FIG. 120 shows the outline of the pitch of the waveform shown in FIG. 119 obtained by the proposed method;
in FIG. 121 shows the oscillogram of the sentence “We didn’t see such a net”, uttered by a VBG male announcer with pause marks at the intervals of the bows of explosive deaf phonemes / t / and / c / in the word “such,” the signal is limited to 5 samples;
in FIG. 122 shows the outline of the pitch of the waveform shown in FIG. 121 obtained by the proposed method;
in FIG. 123 shows the oscillogram of the sentence “We didn’t see such a net”, pronounced by a VBG male announcer with pause marks at the intervals of the bows of explosive deaf phonemes / t / and / c / in the word “such”, the signal is limited to 1 count;
in FIG. 124 shows the outline of the pitch of the waveform shown in FIG. 123 obtained by the proposed method;
in FIG. 125 is a zoomed-in waveform of the bow / d / for the word “not seen” shown in FIG. 101;
in FIG. 126 shows the oscillogram of the sentence "We did not see such a net", uttered by the announcer male AIO;
in FIG. 127 is a waveform of the resynthesized signal of FIG. 126 using the pitch outline obtained by the proposed method and parameters determined by the correlation method on the pitch period without multiplying by the window;
in FIG. 128 is a waveform of the resynthesized signal of FIG. 126 using the outline of the fundamental tone obtained by the proposed method and the parameters determined by the correlation method on the period of the fundamental tone with the multiplication of the signal by an exponential window;
in FIG. 129 is a picture of the formant paths for the signal shown in FIG. 126;
in FIG. 130 is a picture of the formant paths for the signal shown in FIG. 127;
in FIG. 131 is a picture of the formant paths for the signal shown in FIG. 128;
in FIG. 132 shows the waveform of the word "four," pronounced by the male AIO announcer;
in FIG. 133 shows a master pitch reference loop obtained manually for the waveform shown in FIG. 132;
in FIG. 134 is a waveform diagram of the signal of FIG. 132 together with white noise at a signal-to-noise ratio S / N = 5 dB;
in FIG. 135 is a waveform diagram of the signal shown in FIG. 132, smoothed by a secant window;
in FIG. 136 shows the pitch profile obtained by the proposed method from the smoothed signal shown in FIG. 135; in FIG. 137 shows the time function of the secant window;
in FIG. 138 is a waveform of a noise signal shown in FIG. 134 smoothed by a secant window;
in FIG. 139 shows the pitch profile obtained by the proposed method from the smoothed signal shown in FIG. 138, with the formation of the TON / NE_TON sign synchronous with the fundamental tone;
in FIG. 140 shows the pitch profile obtained by the proposed method from the smoothed signal shown in FIG. 138, with the formation of the sign TON / NOT_TON; asynchronous with the fundamental tone;
in FIG. 141 shows the outline of the pitch of the waveform shown in FIG. 134 obtained by the peak method;
in FIG. 142 shows the outline of the pitch of the waveform shown in FIG. 134 obtained by the cepstral method;
in FIG. 143 shows the outline of the pitch of the waveform shown in FIG. 134 obtained by the Rabiner-Gold method;
in FIG. 144 shows the outline of the pitch of the waveform shown in FIG. 134 obtained by the filter method;
in FIG. 145 shows the outline of the pitch of the waveform shown in FIG. 134 obtained by the LLK method;
in FIG. 146 shows the waveform of the signal shown in FIG. 132, after being limited in the frequency band 300 - 3300 Hz (telephone channel);
in FIG. 147 shows the waveform of the signal obtained after smoothing by the secant window of the signal shown in FIG. 146;
in FIG. 148 shows the pitch profile obtained by the proposed method from the smoothed signal shown in FIG. 147, when the decision TON / NOT_TON is synchronized with the fundamental tone;
in FIG. 149 shows the outline of the pitch of the waveform shown in FIG. 146 obtained by the peak method;
in FIG. 150 shows the outline of the pitch of the waveform shown in FIG. 146 obtained by the cepstral method;
in FIG. 151 shows the outline of the pitch of the waveform shown in FIG. 146 obtained by the Rabiner-Gold method;
in FIG. 152 shows the outline of the pitch of the waveform shown in FIG. 146 obtained by the filter method;
in FIG. 153 shows the outline of the pitch of the waveform shown in FIG. 146 obtained by the autocorrelation method;
in FIG. 154 shows the outline of the pitch of the waveform shown in FIG. 146 obtained by the LLK method;
The operation of the algorithm according to the proposed method is illustrated in FIG. 1. The list of operations of the proposed method is shown in FIG. 2. The original signal is passed through a low-pass filter (low-pass filter) with a cutoff frequency in the range from 600 to 1500 Hz (operation 1). The low-pass filter increases the reliability of the allocation of the fundamental tone, although for a number of votes it is not mandatory.

 Positive values of the original signal are logarithmized (step 2). Assuming that the decaying signal, by comparison with which a decision is made to search for the next maximum, is exponential, the operation of logarithmizing the input signal allows the descending signal to be generated in the form of a linearly decreasing signal. The transition to the linear form of the falling signal allows us to simplify the procedure for changing the parameters of this signal in the process of its adaptation to the original input signal. The base of the logarithm is chosen so as to ensure the normalization of the prologarithmic signal within an acceptable range of values of signal processing devices. The logarithm of the positive values of the original signal is shown in FIG. 1 solid thick lines.

In step 3, the first maximum of the original signal, which in FIG. 1 is designated as A ₁ . This maximum corresponds to the marker M ₁ on the time axis, indicating the position of the maximum A ₁ in time.

In the next step 4, the maximum search blocking time is estimated by the formula
T _bl = Q * T ₀ ,
where T _bl - decision blocking time,
Q is a coefficient less than unity and determines what proportion of the period of the fundamental tone should be the blocking time,
T ₀ - period of the fundamental tone, estimated at the previous stages of the algorithm.

When searching for the first maximum, T ₀ is the average period of the fundamental tone of the original signal T _{0 cf} , estimated, for example, as it was done in the LLK method.

Operation 5. Assess the position and maximum value of the falling signal. The peak position of the falling signal is found as
B _k = M _k + T _bl ^k ,
where B _k - the position in time of the maximum of the falling signal at the k-th period of the fundamental tone,
M _k - mark of the k-th period of the fundamental tone,
T _bl ^k - blocking time of the k-th period of the fundamental tone.

The maxima of the falling signal are determined by the formula
G _max ^k = A _k + D,
where G _max ^k - the maximums of the falling signal on the k-th period of the fundamental tone,
A _k - the maximum value of the signal corresponding to the brand M _k on the k-th period of the fundamental tone,
D is the value of the increment of the maximum of the falling signal, providing reliable selection of the subsequent main maximum of the original signal.

 The increment value D is determined from the statistics of the initial signal at the stage of preliminary experiments.

Step 6. Determine the slope of the falling signal. To do this, first find the slope in the plot T _bl , assuming a linear increase in the signal from A _k to G _k :
R _k = D / T _bl ^k ,
where R _k is the slope of the linear increase in the interval T _bl ^{k of the} kth period of the fundamental tone.

The slope of the falling signal in the reduction section is determined as
L _k = R _k * A _s ,
where L _{k the} slope of the falling signal in the area of decline of the k-th period of the fundamental tone,
A _s is the asymmetry coefficient, determined on the basis of statistics of measurements of the period of the fundamental tone in the course of preliminary experiments.

Operation 7. Generate a falling signal in accordance with the formula
G _k (t) = G _max ^k - L _k (t - B _k ),
where G _k (t) is the value of the falling signal at time t,
t is the current time,
B _k - the position in time of the maximum of the falling signal at the k-th period of the fundamental tone.

Step 8. At each step of determining the falling signal, check the condition
G _k (t) ≥ log _a (s (t)),
where s (t) is the original signal,
and - the base of the logarithm.

If this condition is met, then continue to determine G _k (t) for subsequent values of t. If the specified condition is violated, then proceed to the search for the next maximum. In FIG. 1 point corresponding to violation of the specified condition for the first period of the fundamental tone is designated as C ₁ , for the second - C ₂ .

Operation 9. After assessing the violation of the above conditions, they proceed to search for the next main maximum of the original signal. For the second pitch period, this point in FIG. 1 is designated as A ₂ , it corresponds to the temporary mark M ₂ . By difference
M _{k + 1} - M _k = T ₀ ^k
determine the value of the k-th period of the fundamental tone. If the value of T ₀ ^k does not go beyond the permissible limits of changes in the period of the fundamental tone, i.e. if
T $\genfrac{}{}{0ex}{}{\text{k-1}}{\text{0}}$ - Δ ^* T $\genfrac{}{}{0ex}{}{\text{k-1}}{\text{0}}$ ≤ T $\genfrac{}{}{0ex}{}{\text{k}}{}$$\genfrac{}{}{0ex}{}{}{\text{0}}$ ≤ T $\genfrac{}{}{0ex}{}{\text{k}}{}$$\genfrac{}{}{0ex}{}{}{\text{0}}$ + Δ ^* T $\genfrac{}{}{0ex}{}{\text{k-1}}{\text{0}}$ ,
where Δ is the allowable relative change in the period of the fundamental tone, determined on the basis of statistics of the initial signal during preliminary experiments,
then when determining the duration of the blocking time for the (k + 1) -th period of the fundamental tone T _bl ^{k + 1, the} value of T _{0 is} replaced by T ₀ ^k , otherwise the value of T ₀ in the formula for determining T _{bl is} left unchanged.

 Then go to step 4, and the process of evaluating a new period of the fundamental tone is repeated.

 In FIG. Figures 3 to 143 show the results of a comparative analysis of the allocation of the fundamental tone by the proposed method and the considered analogues (peak method, filter method, Rabiner-Gold method, cepstral method, autocorrelation method, LLK method). Analogue methods were implemented on the STC-D 106.1 installation using the SIS 4.0 software system developed and manufactured by the Center for Speech Technologies company in St. Petersburg [Methodological recommendations for the practical use of the SIS program when working with speech signals / Speech Technology Center. - S. Petersburg, 1997]. The proposed method is implemented in the form of programs written in Turbo-Pascal, SI and SI ++.

In all experiments, a quantization frequency of 10,000 Hz was used. All analog methods work as follows:
1. Cutting off low-energy signal sections. For this, the average value of the signal amplitude is calculated on each signal analysis frame specified by the duration and the calculated value is compared with a threshold set by the user. Plots, the average value of the amplitude on which does not exceed the threshold, are taken for pauses.

 2. Cutting off high-frequency noise sections of the signal. To do this, on each analysis frame, the average value of the frequency of zero crossing is calculated and the calculated value is compared with the threshold set by the user. Sites, the average value of the frequency of zero crossing at which exceeds a predetermined threshold, is taken as noise.

 3. Determination of the value of the period of the fundamental tone.

 4. Check the degree of periodicity of the signal for a given period of the fundamental tone.

 5. Decision making Tone / Ne_ton.

 6. Go to the next frame.

 The proposed method selects the main tone of the signal without first dividing the signal into sections with low and high energy, into sections of noise and sections of speech. The TONE / NOT_TONE decision can be made after selecting the pitch outline, which allows you to additionally take into account the information available in the pitch circuit and thereby improve the reliability of the TONE / NOT_TONE decision.

 In the peak pitch extraction method implemented in the SIS program, frames with a duration of, for example, 20 ms are extracted from the original signal. The beginning of each subsequent analysis frame is shifted by the current pitch period compared to the previous frame. On the first vocalized segment of speech, they look for the value of the period of the fundamental tone and the beginning of each period at the maximum amplitude. Next, in the area spaced from the maximum by the current value of the period of the fundamental tone, the next maximum is sought. The distance between the maxima is taken as the new value of the fundamental tone at this step. For the obtained value of the fundamental tone, the autocorrelation coefficient of the original signal is considered and compared with the threshold. If it exceeds the threshold, then the frame is considered vocalized and the pitch is considered defined, if not, then noise. The method is very sensitive to parameter settings [Methodological recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. - S. Petersburg, 1997. - S. 276].

 In the filter method, the signal is passed through a narrow-band filter before analysis begins. For male voices, the filter passband is 50 - 250 Hz, for women - 70 - 450 Hz. Next, an analysis similar to the peak method is performed. The method is often not workable for signals with a suppressed low frequency range [Methodological recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. - S. Petersburg, 1997. - S. 278].

 A brief description of the Rabiner-Gold algorithm is given above. In this case, too small peaks are discarded in the vicinity of large peaks and the peaks of the temporary signal are estimated in terms of width and amplitude to search for only sufficiently similar peaks. The obtained three current estimates of the duration of the periods of the fundamental tone are considered together for the current, previous and subsequent periods of the fundamental tone. All estimates are compared and for the assessment of the current period of the fundamental tone, take the one that is most often found in this set of assessments of the fundamental tone. The algorithm operation scheme corresponds to the peak method scheme [Methodological recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. - S. Petersburg, 1997. - S. 278].

 In the cepstral method for isolating the fundamental tone to determine the cepstrum, the inverse Fourier transform of the complex logarithm of the signal power spectrum in the analysis frame is performed based on the FFT algorithm. The duration of the analyzed frame should exceed the duration of at least the two longest periods of the fundamental tone and be a multiple of a power of two, which for male voices is usually 51.2 ms and for female 25.6 ms. The original signal is multiplied by the Hann or Haming window. The shift of the analysis frame is set within 1/16 - 1/4 of the length of the analysis frame. To increase the reliability of the allocation of the fundamental tone can be used underlining cepstrum in the field of large values. For each analyzed frame, the obtained cepstrum is examined in order to search for a peak in the region of possible values of the fundamental tone. Unlike the peak method, the pitch is determined on average for the analysis frame, which leads to a smoothing of the trajectory of the pitch. The rest of the scheme of the algorithm corresponds to the peak method [Methodological recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. - S. Petersburg, 1997. - S. 276].

 In the autocorrelation method for isolating the fundamental tone in the region of the estimated value of the period of the fundamental tone, the autocorrelation function is calculated and its maximum is sought. The maximum value is compared with the threshold and a decision is made about the actual vocalization of the segment. In the case of a positive decision, the position of the maximum of the autocorrelation function is taken as the value of the period of the fundamental tone. The rest of the scheme of the algorithm corresponds to the peak method [Methodological recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. - S. Petersburg, 1997. - S. 277].

 When isolating the fundamental tone using the LLK method, the autocorrelation with the 20 ms window is calculated for the voiced sections. Using 15 successively calculated autocorrelation values, an assumption is made about the magnitude of the pitch period. Knowing the estimated value of the period of the fundamental tone, the peak emitter of the fundamental tone forms the final result. Methodological recommendations for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. - S. Petersburg, 1997. - S. 279].

 When isolating the fundamental tone using analog methods, the following parameters for signal and noise were used in the SIS program.

 Parameters for the signal.

 a). Frame size - the required duration of the analysis frame in samples.

 b) The initial frequency of the fundamental tone (F) is the value of the frequency of the fundamental tone at each transition from the noise section to the voiced one. The first value of the frequency of the fundamental tone at each transition from the noise section to the vocalized one is searched for in the interval from F-ID to F + ID, where ID is the initial smoothness value.

 in). Smoothness of the melodic curve (in counts) - sets the maximum possible interval for the change in the current period of the fundamental tone in relation to the previous period.

 d). The initial smoothness of the melodic curve (in samples) - sets the maximum possible interval in which the first value of the fundamental tone is searched for at each transition from the noise section to the vocalized one.

 e). The absolute minimum frequency of the fundamental tone (Hz) is the value of the fundamental tone, below which the fundamental tone is not sought.

 e). The absolute maximum frequency of the fundamental tone is the value of the fundamental tone, above which the fundamental tone is not sought.

 Parameters for noise and pause.

 a). Frame size for analysis of noise and pauses - determines the length of the analysis frame when deciding (by energy and frequency of zero intersections) whether the frame is voiced, noise or pause.

 b) Frame shift for noise and pause analysis - sets the size of the frame shift in time when classifying the frame as noisy or pause.

 in). Threshold for determining the pause - determines the threshold value of the amplitude. A frame with an amplitude below the threshold is referred to as a pause.

 d). Threshold for determining the frequency of zero crossing for determining noise (Hz). The signal in the analysis frame with the frequency of crossing the zero level above the threshold is considered noise.

 e). The severity parameter of autocorrelation. When calculating the pitch for each pitch value obtained, the autocorrelation coefficient is calculated and compared with a threshold. If it exceeds the threshold, then the frame is classified as vocalized and the pitch is considered defined, if not, then the frame is considered noise.

 The indicated parameters require a preliminary analysis of the initial signal, and in many cases the shape of the resulting fundamental tone circuit is critical to the choice of parameters for noise and pause.

 In FIG. 3 shows an oscillogram of the sentence “We didn’t see such a net”, uttered by a male announcer (VBG), in FIG. 4 reference pitch outline for the specified phrase. The reference circuit is obtained manually by the waveform. For the beginning of the period of the fundamental tone, the moment of crossing the zero with the main lobe corresponding to the beginning of the excitation of the vocal tract was taken. This approach allows you to more accurately determine the beginning of the period of the fundamental tone, because often the main maximum on the period of the fundamental tone is not very clearly expressed.

 In FIG. 5–10 show the contours of the fundamental tone identified by the peak method, the filter method, the Rabiner – Gold method, the cepstral method, the autocorrelation method, and the LLK method, respectively, with a standard set of parameters for men. The standard set of parameters for the male voice is given in table. 1. It can be seen that only the LLK method ensured a satisfactory quality of distinguishing the trajectory of the fundamental tone.

In FIG. 11 shows the trajectory of the fundamental tone obtained by the proposed method with the following parameters: cutoff frequency of the low-pass filter 1500 Hz, initial frequency of the fundamental tone 120 Hz, permissible change in the period of the fundamental tone Δ = 0.5, the increment of the decaying signal relative to the selected maximum D = 200 at provided that the signal is normalized to a level of no more than 2048 samples, the blocking coefficient is Q = 0.5, the asymmetry coefficient is A _s = 1.8, the maximum permissible fundamental frequency is F _max = 500 Hz. In subsequent experiments, these parameters remained unchanged, with the exception of some cases, which will be indicated separately. Pause highlighting is done manually. From FIG. 11 it can be seen that the proposed method provides high-quality selection of the trajectory of the fundamental tone for the male voice, not worse than in the LLK method.

 In FIG. 12 shows the same waveform as in FIG. 3, but with the specified position of the marks corresponding to pauses.

 In FIG. Figures 13 - 19 show the contours of the fundamental tone of the indicated sentence, identified after adjusting the parameters by the peak method, the filter method, the Rabiner-Gold method, the cepstral method, the autocorrelation method, the LLK method and the proposed method, respectively. The parameters obtained after adjustment are shown in Table 2. The parameters of the proposed method did not change.

 After adjusting the parameters, a significant increase in the quality of the selected contours was achieved. The best allocation of contours among analog methods was obtained for the cepstral method and the LLK method. The quality of the outline of the fundamental tone by the proposed method is not worse than in the analogue methods. None of the analogue methods provided the correct allocation of the trajectory of the fundamental tone for the short vowel / a /, located between the bows / t / and / c / in the word "such".

 In FIG. Figure 20 shows the utterance of the sentence “We didn’t see such a net” by the announcer woman (IHF). In FIG. 21 is a reference outline of the fundamental tone obtained manually by the waveform. In FIG. 22 - 28 show the contours of the fundamental tone of the sentence, identified after adjusting the parameters by the peak method, the filter method, the Rabiner-Gold method, the cepstral method, the autocorrelation method, the LLK method and the proposed method, respectively. The parameters obtained after adjustment are shown in table. 3. The parameters of the proposed method did not change except the initial frequency of the fundamental tone, which is taken equal to 240 Hz. The conclusions regarding the quality of emphasizing the fundamental tone by analogue methods and the proposed method are similar to the conclusions drawn for the male voice. In FIG. Figure 29 shows an oscillogram of the word "three," pronounced in isolation by a male announcer (VBG). In FIG. 30 - 36 show the contours of the fundamental tone of the specified word, highlighted after adjusting the parameters by the peak method, the filter method, the Rabiner-Gold method, the cepstral method, the autocorrelation method, the LLK method and the proposed method, respectively. The parameters obtained after adjustment are shown in table. 4. The parameters of the proposed method correspond to the parameters used to highlight the main tone of the sentence "We did not see such a net", uttered by the announcer VBG.

 The analysis obtained in FIG. 30 - FIG. 36 circuits shows that none of the analogue methods provided the correct allocation of the pitch circuit at the transition from the bow / t / to the bell / p /. To prove the correspondence of the pitch contour obtained by the proposed method to the true pitch contour in FIG. 37, FIG. 39 and FIG. 41 shows, in an enlarged time scale, the oscillograms of individual sections of the word “three” filtered by a low-pass filter with a cutoff frequency of 1500 Hz, and in FIG. 38, FIG. 40 and FIG. 42 - the corresponding contours of the fundamental tone obtained by the proposed method. It can be seen from these figures that the proposed method ensured the correct marking at the beginning of the excitation of the vocal tract, both in transitional and stationary pronunciation areas.

 In FIG. 43 the pronunciation of the sentence “We didn’t see such a net” is given by the VBG speaker with added white noise at a signal to noise ratio S / N = 40 dB. The ratio was determined in the signal sections with maximum energy. In FIG. 44-50 show the contours of the fundamental tone of the indicated pronunciation with noise at S / N = 40 dB obtained by the peak method, filter method, Rabiner-Gold method, cepstral method, autocorrelation method, LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for pronunciation without noise are shown in table. 5. The parameters of analog methods obtained after adjustment for pronunciation with added noise at S / N = 40 dB are shown in Table. 6 The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and did not change. The conclusions regarding the quality of emphasizing the fundamental tone by analogue methods and the proposed method are similar to the conclusions drawn for pronouncing without noise in the analysis of the oscillogram shown in FIG. 3.

 In FIG. Figure 51 shows the pronunciation of the sentence “We didn’t see such a net” by the VBG speaker with added white noise at a signal to noise ratio S / N = 30 dB. In FIG. 52 - 58 show the contours of the fundamental tone of the indicated pronunciation with noise at S / N = 30 dB obtained by the peak method, filter method, Rabiner-Gold method, cepstral method, autocorrelation method, LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for pronunciation with noise at S / N = 30 dB are shown in Table. 7. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and did not change.

 We can note the appearance of errors in determining the fundamental tone by all methods without exception. The worst results were shown by the autocorrelation method (Fig. 57). The cepstral method worked best (Fig. 55), the LLK method (Fig. 57) and the proposed method (Fig. 58). And in this case, you can give preference to the proposed method. His mistakes are mainly located in the phoneme / a / between the two bows / t / and / c / of the word "such". The specified section by analogous methods in all cases was processed with errors.

 In FIG. 59 the pronunciation of the sentence “We didn’t see such a net” is given by the VBG speaker with added white noise at a signal to noise ratio S / N = 20 dB. In FIG. 60 - 66 shows the contours of the fundamental tone of the indicated pronunciation with noise at S / N = 20 dB obtained by the peak method, filter method, Rabiner-Gold method, cepstral method, autocorrelation method, LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for pronouncing with noise at S / N = 20 dB are shown in Table. 8. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and did not change.

 The number of errors in all methods compared with the pronunciation at S / N = 30 dB increased, but not significantly. Again, the worst results were shown by the autocorrelation method (Fig. 64). From FIG. 66 it follows that even with S / N = 20 dB, one can give preference to the proposed method, although a small number of additional errors can be noted on the bow interval / d / in the word "net".

 In FIG. 67 shows the pronunciation of the sentence “We didn’t see such a net” by the VBG speaker with added white noise at a signal to noise ratio S / N = 10 dB. In FIG. 68 - 74 show the contours of the fundamental tone of the indicated pronunciation with noise at S / N = 10 dB obtained by the peak method, filter method, Rabiner-Gold method, cepstral method, autocorrelation method, LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for pronouncing with noise at S / N = 10 dB are shown in Table. 9. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and did not change except the cutoff frequency of the low-pass filter, which is taken to be 800 Hz.

 The number of errors in all methods compared with pronunciation in the absence of noise increased. In all analogue methods, the pronunciation ending section is lost, errors appeared on the phoneme / n / in the word "net". In the peak, filter, autocorrelation and LLK methods, there are errors in the transition from / and / to / m / in the combination of the words “we did not see”. In all analogue methods, the section of short / a / between closure / t / and / c / was processed with large errors. The irregularity of the outline of the fundamental tone obtained by the proposed method has increased, the number of errors in the section of the bow / d / from the word "not seen" has also increased. However, only the proposed method as a whole correctly received the trajectory of the fundamental tone, including at the end of pronunciation and at the short / a / between the bows / t / and / k /.

 In FIG. 75 the pronunciation of the sentence “We didn’t see such a net” is given by the VBG speaker with added white noise at a signal to noise ratio S / N = 5 dB. In FIG. 76 - 82 shows the contours of the fundamental tone of the specified pronunciation with noise at S / N = 5 dB obtained by the peak method, the filter method, the Rabiner-Gold method, the cepstral method, the autocorrelation method, the LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for pronunciation with noise at S / N = 5 dB are shown in Table. 10. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and did not change except the cutoff frequency of the low-pass filter, which is taken to be 800 Hz.

 The number of errors in all methods compared with pronunciation in the absence of noise has increased significantly. In all analog methods, the areas of the beginning and end of pronunciation are lost, there are numerous errors in the transitions between phonemes.

 The number of errors in the allocation of the fundamental tone by the proposed method has also increased, it can be noted the transition to half the frequency of the fundamental tone at the beginning of pronunciation, increased ruggedness and an increased number of errors at the end of pronunciation. However, the proposed method as a whole ensured the correct tracking of the trajectory of the fundamental tone throughout the entire pronunciation.

 In FIG. 83 the pronunciation of the sentence “We didn’t see such a net” is given by the VBG speaker with added white noise at a signal to noise ratio S / N = 0 dB. In FIG. 84 - 87 shows the contours of the fundamental tone of the indicated pronunciation with noise at S / N = 0 dB obtained by the peak method, the filter method, the Rabiner-Gold method and the proposed method, respectively. The cepstral method, the autocorrelation method, and the LLK method were completely inoperative. The parameters of the analog methods obtained after adjustment for pronouncing with noise at S / N = 0 dB are shown in Table. 11. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and did not change except the cutoff frequency of the low-pass filter, which is taken to be 800 Hz.

 The number of errors in the allocation of the fundamental tone by the proposed method compared with the case of the ratio S / N = 5 dB increased especially at the end of pronunciation. However, the proposed method, in general, despite such a low S / N ratio, provided acceptable tracking of the trajectory of the fundamental tone throughout the pronunciation, while analogue methods were practically inoperative.

 To prove the correct operation of the proposed method with the ratio S / N = 0 dB in FIG. 88 shows the original signal after a low-pass filter with a cutoff frequency of 800 Hz, FIG. 89 is the original signal with noise at S / N = 0 dB, in FIG. 90 is a pitch profile obtained for a noisy signal at S / N = 0 dB, in FIG. 91-100 are waveforms of the signal shown in FIG. 88, and the pitch outline shown in FIG. 90, enlarged over time.

 In FIG. 91 and 92 show the oscillogram of the transition from the bow / d / to the vowel / a / and the stationary part / a / from the word "not seen" and the corresponding pitch circuit obtained by the proposed method using a noisy signal. In all cases, the marks correspond to the moments of excitation of the vocal tract.

 In FIG. 93 and 94 show the oscillogram of the transition from to the vowel / a / to the sonor / l / further to the vowel / and / from the word "not seen" and the corresponding pitch circuit obtained by the proposed method for a noisy signal. It is possible to note errors in the marking on / l / having a relatively low level of intensity. However, the transition to / and / is again processed correctly.

 In FIG. Figures 95 and 96 show the oscillogram of the transition from the sonoric / m / to the vowel / s / from the word "we" and the corresponding pitch circuit obtained by the proposed method using a noisy signal. Two errors of marking on / m / can be noted, which gave rise to the pitch of the main tone. From FIG. 95 it can be seen that even on a non-noisy signal at the fundamental tone period in / m / there are repeated maxima comparable in amplitude to the main ones, which led to failures in the signal with noise. However, in the transition to / s / marks are affixed correctly.

 In FIG. 97 and 98 show the oscillogram of the vowel / a / located between the bows / t / and / c / in the word "such" and the corresponding pitch circuit obtained by the proposed method using a noisy signal. In all cases, the marks on the vowel / a / are affixed correctly. In FIG. 99 and 100 show the oscillogram of the transition from the vowel / o / to the bow / d / in the word "net" and the corresponding pitch circuit obtained by the proposed method for a noisy signal. It can be seen that despite the fact that the marks do not always correspond to the moments of excitation of the vocal tract, the frequency of the fundamental tone in the considered pronunciation section is preserved.

 With a further decrease in the S / N ratio, the proposed method turned out to be inoperative. In general, it can be argued that the proposed method provides the allocation of the trajectory of the fundamental tone with a S / N ratio of at least 5 dB lower compared to analog methods.

 In FIG. 101 and 102 show the oscillogram of the speaker’s utterance by the VBG sentence “We didn’t see such a net” and the pitch outline obtained by the proposed method, respectively.

 This circuit in the experiments below is taken as a reference. The maximum signal level in FIG. 101 is equal to 819 samples.

 In FIG. 103 is a waveform of a signal obtained from the signal shown in FIG. 101, by limiting the latter to 100 samples. In FIG. 104 - 110 shows the contours of the fundamental tone of the indicated pronunciation with a limit of 100 samples obtained by the peak method, the filter method, the Rabiner-Gold method, the cepstral method, the autocorrelation method, the LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for the signal limited at the level of 100 samples are shown in Table. 12. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and remained unchanged. From FIG. 104 - 110 it follows that all the methods presented for the signal limited at the level of 100 samples, provided a satisfactory quality of the selection of the fundamental tone.

 In FIG. 111 is a waveform diagram of a signal obtained from the signal shown in FIG. 101, by limiting the latter to 50 samples. In FIG. 112 - 118 shows the contours of the fundamental tone of the specified pronunciation with a limit of 50 samples obtained by the peak method, filter method, Rabiner-Gold method, cepstral method, autocorrelation method, LLK method and the proposed method, respectively. The parameters of the analog methods obtained after adjustment for the signal limited at the level of 50 samples are shown in Table. 13. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and remained unchanged. From FIG. 112 - 117 it follows that all the methods presented, except for the autocorrelation, provided a satisfactory quality of the selection of the fundamental tone. It is possible to note the appearance of individual errors on all methods except the cepstral one. The proposed method (Fig. 118) has kept its circuit unchanged.

 With a further decrease in the level of restriction, analogue methods turned out to be inoperative. The proposed method, nevertheless, ensured a reliable selection of the fundamental tone contour up to the limiting limit level of 1 count. In FIG. 119 -124 shows the waveforms of the signals and the corresponding pitch paths obtained by the proposed method for the levels of limitation of 10 samples, 5 samples and 1 sample. To illustrate the presence of periodicity of the fundamental tone in the section of the bow / d / from the word “not seen” in FIG. 125 shows the original signal in a zoomed-in time scale. In all cases, the parameters of the proposed method remained unchanged.

 The proposed method relates to local methods, i.e. it determines the moment of excitement at each period of the fundamental tone. The indicated property of the proposed method allows the analysis of the speech signal at the period of the fundamental tone and to obtain additional useful effects.

 It is known [J. D. Markel, A.X. Gray. Linear prediction of speech. - M .: Svyaz, 1980. - 308 pp.] That the correlation method of linear prediction does not ensure the stability of the solution in the process of calculating linear prediction coefficients. In FIG. Figure 126 shows an oscillogram for pronouncing the sentence "We didn’t see such a net", uttered by an announcer male AIO. On the indicated waveform by the proposed method, stamps of the beginnings of excitations of the speech tract are affixed. The parameters of the proposed pitch method are completely identical to the parameters used to extract the signal shown in FIG. 3 (VBG speaker). After marking, synchronously with the fundamental tone, an analysis was performed by the correlation method of linear prediction of the 12th order. The number of samples in each analysis frame corresponds to the length of one period of the fundamental tone (from one brand to the next). A re-synthesized signal of said utterance is shown in FIG. 127. In FIG. 127 the moments of unstable decisions are clearly visible at the moments of transition from audible to vocalized sounds.

 The proposed method allows to eliminate this drawback. For this purpose, each period of the fundamental tone between two consecutive marks is multiplied by an exponentially decaying window. The exponent decay rate is chosen so as to ensure the stability of the solution of the correlation method. At the stage of resynthesis, in order to restore the signal obtained on the basis of linear prediction coefficients, the resynthesized signal is multiplied by a rising exponent with an exponent corresponding to the stage of analysis. An example of a signal thus synthesized is shown in FIG. 128. From FIG. 128 it follows that the proposed signal processing method ensures the stability of the solution when using the correlation linear prediction method.

 Experiments on listening to a re-synthesized signal have shown that it has good quality and intelligibility. In FIG. 129, 130, and 131 show the paths of the formants of the original signal resynthesized without multiplying by the exponent of the signal and resynthesized using multiplication by the exponential of the signal, respectively. It can be noted that for most of the pronunciation, the formant structure as a whole is preserved for both types of resynthesized signal.

 A signal mixed with noise largely loses its frequency associated with the fundamental tone. To emphasize the indicated periodicity, it is proposed to use the moving average method to smooth the signal multiplied by the time window. The sch (t) window was used as a temporary window. In FIG. 135 - 138 show the waveform of the word "four" spoken by an AIO male announcer; a reference outline of the fundamental tone, for the specified word, constructed manually; waveform of the signal of this word with added noise at S / N = 5 dB; an oscillogram of a smoothed secant window and a normalized signal without noise; the main tone circuit constructed by a smoothed signal without noise by the proposed method; the time function of the secant window and the signal with noise smoothed by the secant window. The width of the secant window at the level of 0.05 from the maximum is taken equal to 0.67 from the average period of the fundamental tone. The parameters of the proposed method remained unchanged.

 From a comparison of FIG. 132 and 135, FIG. 134 and 138 it can be seen that smoothing by the secant window provides underlining of the frequency associated with the fundamental tone while simplifying the shape of the processed signal. The pitch profile obtained for a smoothed signal without noise (Fig. 136) corresponds to the reference contour (Fig. 133).

 The proposed method affixes marks at time points corresponding to the moments of excitation of the vocal tract. The placement of these brands allows you to analyze the characteristics of the speech signal synchronously with the fundamental tone within the same period of the fundamental tone. The parameters of the speech-forming tract on one period of the fundamental tone do not change significantly, which allows to increase the accuracy and reliability of the measurement of these parameters. Experiments have been conducted to evaluate the accuracy of the TON / NE_TON decision based on the behavior of three signs: 1) energy, 2) the frequency of zero crossing, and 3) the first linear prediction coefficient. The decision was made synchronously and asynchronously with the main tone. If a decision was made in synchronism with the fundamental tone, the indicated parameters were measured on a signal between two consecutive marks and, after exceeding the experimentally established thresholds, together with three signs, they made a decision about the presence or absence of vocalization for a given time interval. In the case of asynchronous analysis, the TON / NE_TON decision was taken on the analysis frame with a duration of 20 ms. Next, the analysis frame was shifted by 1 ms and the decision procedure TON / NE_TON was again performed. The contours of the fundamental tone constructed by the proposed method using a smoothed, noisy signal (Fig. 138) in the case of synchronous and asynchronous TON / NOT_TON decisions are shown in FIG. 139 and 140, respectively. From a comparison of these circuits with the reference circuit of the fundamental tone (Fig. 133) it is seen that the decision TON / NOT_THON synchronous with the fundamental tone provides fewer errors.

In FIG. 141 to 145 show the outline of the pitch of the noise signal shown in FIG. 134 obtained by the peak method, the cepstral method, the Rabiner – Gold method, the filter method, and the LLK method, respectively. The parameters of the analog methods obtained after adjustment for the signal with noise at S / N = 5 dB are shown in Table. 14. The parameters of the proposed method correspond to the pronunciation of the male voice (Fig. 3) and remained unchanged except for the asymmetry coefficient A _s , which was increased from 1.8 to 2.0. From FIG. 141 - 145 it follows that all the methods presented provided the worst quality of the TON / NE_TON decision in comparison with the TON / NE_TON decision synchronous with the main tone based on the proposed method. The autocorrelation method was generally inoperative.

 In FIG. 146 shows a telephone channel signal obtained by passing the signal shown in FIG. 132, through a band-pass filter with cutoff frequencies of 300 - 3300 Hz. The eighth-order Butterworth filter is used as a band-pass filter. From FIG. 146 it can be seen that suppressing the first harmonic range of the fundamental tone suppresses fundamental information in the signal.

 The operation of smoothing with a secant window of a telephone signal allows you to restore information about the fundamental tone. As for a noisy signal, smoothing with a secant window is performed after the filtering operation before the logarithm operation. The width of the secant window at the level of 0.05 from the maximum is taken equal to 0.67 from the average period of the fundamental tone. The telephone signal normalized and smoothed by the secant window is shown in FIG. 147. From FIG. 147 it can be seen that smoothing with a secant window does indeed allow a substantial recovery of information about the fundamental tone.

 The pitch outline obtained from the telephone signal smoothed by the secant window (FIG. 147) is shown in FIG. 148. The decision TONE / NOT_TONE is executed synchronously with the fundamental tone. The parameters of the proposed method remained unchanged, except for the asymmetry coefficient, which was taken equal to 2.6. From a comparison with the reference circuit (Fig. 133), it can be seen that in this case, the proposed method provided a satisfactory quality of the allocation of the fundamental tone, although errors at the beginning and at the end of vocalization can be noted.

 In FIG. 149 to 154 are the pitch profiles for the telephone signal shown in FIG. 146 obtained by the peak method, cepstral method, Rabiner-Gold method, filter method, autocorrelation method, and LLK method, respectively. The parameters of analog methods for a telephone signal are shown in Table. 15. From a comparison with the reference (FIG. 133) of the circuits shown in FIG. 149 - 151, it can be seen that the peak method, the cepstral method, and the Rabiner-Gold method provided a satisfactory quality of the extraction of the fundamental tone by the telephone signal, although some features of the circuit were lost. And under these conditions, the circuit obtained by the proposed method is preferred. The filter method, the autocorrelation method, and the LLK method (Fig. 152 - 154) showed a rather large number of errors.

LITERATURE
1. M. A. Sapozhkov. Speech signal in cybernetics and communication. - M .: Svyazizdat, 1963 .-- 472 p.

 2. B. Gold, L. Rabiner. Parallel processing techniques for estimating pitch period of speech in the time domain // J. Acoustic Soc. Am., 1969 - 46. - N 2 (Pt. 2). - P. 442 - 448.

 3. A.M. Noll. Short-Time spectrum and "Cepstrum" techniques for vocal-pitch detection // J. Acoustic. Soc. Am. - 36. - N 2, 1955.

 4. A.A. Pies. Device for automatically highlighting the fundamental tone. - A.S. N 129739 USSR. - Prior. from 06/08/58. - NKI 21e 1/20 42e.

 5. Guidelines for the practical use of the SIS program when working with speech signals / Center for Speech Technologies. S. Petersburg. - S. Petersburg, 1997 .-- 394 p.

 6. L.O. Dolansky. Instantaneous pitch period indicator // J. Acoust. Sos. Am. - 27. - N 11, 1955. - P. 67 - 72 (prototype).

Claims (7)

 1. The way to select the fundamental tone, which consists in the fact that the main maxima of the time function of the original signal are extracted, after the maximum is selected, the search for the maximum is blocked during the blocking interval, after the blocking interval, a decaying signal is generated, decreasing relative to the highlighted maximum, characterized in that the decaying signal is generated with a maximum value that at the end of the blocking time exceeds the absolute maximum of the original signal in the analyzed period of the fundamental tone.  2. The method according to claim 1, characterized in that the original signal is pre-logarithm, and the falling signal is built in the form of a linearly decreasing signal.  3. The method according to claim 1, characterized in that the blocking time and the rate of decrease of the falling signal are adapted to the current period of the fundamental tone.  4. The method according to claim 1, characterized in that the signal is preliminarily limited in frequency by a low-pass filter or a band-pass filter.  5. The method according to claim 1, characterized in that the original signal between the selected main maxima is multiplied by an exponentially decreasing signal.  6. The method according to claim 1, characterized in that before the logarithm operation, the signal is smoothed by a time window, the duration of which is determined by the duration of the period of the fundamental tone.  7. The method according to claim 1, characterized in that the evaluation of the signal parameters is performed on the time interval between the selected main maxima of the original signal.

Images