RU2180974C2 - Process of compression of insulated layers - Google Patents

Abstract

FIELD: digital processing of speech. SUBSTANCE: process is intended for use in systems of speech answer to subscribers of telephone network, in information services, for sounding announcements on transport or in other public places. Process includes division of signal of due word preliminary recorded in on-line storage into sections of equal length, computation of mean modulus of this signal and number of changes of sign in each of them. These data are used to define two images of processed word that describe character of change of its signal by level and instantaneous frequency in time. Decreased capacity of speech information needed for storage and playback is obtained by determination of sections of local stationarity in which level of signal and its instantaneous frequency almost do not change simultaneously, by isolation of section of signal which serves as standard period of basic tone of speech inside each such section, by re-recording of given sections of signal one after another into permanent storage supplying each of them with password containing information on duration of given section of signal, on number of its repetitions during of playback of word and on value of adaptive period of quantization proportional to mean modulus of signal across given section of local stationarity. EFFECT: decreased capacity of information storage used for memorization and playback of required speech information. 4 dwg

Description

 The invention relates to techniques for digital speech processing and can be used in various applications, for example, in voice response systems for telephone network subscribers (answering machines), in help services, for voicing announcements in vehicles, in public places, etc.

 The well-known audio data compression algorithm ISO / MPEG (MUSICAM), using information compression to transmit high-quality sound signals of television programs, as well as digital satellite broadcasting programs [1]. This algorithm is based on the perception of sounds by the human ear, the so-called psychoacoustic effect. It is proved that a person perceives about 10% of the information contained in an audio signal, the remaining 90% are redundant and can not be transmitted over the communication channel. A signal of a certain frequency (tone), acting on a person’s ear, does not allow one to distinguish (mask) other tones that are close to a given frequency and lower in level. In a real sound signal, several masking tones at different frequencies are present simultaneously. Signal components whose levels are below the masking threshold are not perceived by the ear and are redundant.

 The disadvantage of the described audio signal compression algorithm is the difficulty in its implementation. As in frequency vocoders, the compression analyzes the instantaneous energy spectrum of the signal, in this case, using a comb of filters dividing the spectrum into 32 frequency bands. In each of them, an analog-to-digital conversion is performed separately, processing is carried out in succession "frames" of the signal with a duration of 24 ms, with a sampling frequency of samples equal to 48 kHz. Elimination of frequency bands with a signal level below the masking threshold from the transmitted “frame”, in combination with the dynamic distribution of information bits between the remaining bands, allows achieving almost 6-fold compression of the stereo signal spectrum, while maintaining an almost unchanged, very high sound quality.

 The complexity in the implementation of the algorithm [1] is justified by the need to obtain high quality sound of a broadcasting stereo signal, which is not required at all in devices such as answering machines, for help services, etc., where it is enough to ensure high intelligibility and naturalness of speech.

 The closest technical solution (prototype) is the digital signal conversion algorithm for an audio signal using isolated words spoken by an arbitrary speaker as described in [2]. This adaptive algorithm, using the redundancy of the speech signal in the time domain, allows for the recognition of arbitrary voice, regardless of its volume, speech rate and pitch frequency. In contrast to [1], where, as in all known vocoder speech transmission systems, signal compression is achieved by eliminating its frequency redundancy, [2] shows the possibility of using temporal redundancy of a speech signal. This redundancy is manifested in strong correlation relationships, covering up to (10-12) neighboring samples of the speech signal, taken with a sampling frequency of 8 kHz. In turn, the connections between neighboring samples are caused by a sharp unevenness of the power spectrum of the speech signal, which has a maximum in the region of (400-500) Hz and quickly decreases at high frequencies with a speed of 6 to 12 dB per octave.

 Speech redundancy is especially noticeable when pronouncing the so-called voiced (vowels) sounds, which correspond to sections of local stationarity with a length of up to (150-200) ms. Each such section contains dozens of almost the same type of signal segments with a period of the main tone of speech, individual for each voice. For male voices, this period of oscillation of the vocal cords is (5-20) ms, and for high female and children's voices it varies in the range (2-5) ms.

 Pronouncing voiced consonants (such as "b", "c", "g", "d", etc.), as well as "m", "n", is also accompanied by a periodic repetition of segments of the main tone of speech, only their shape changes, amplitude and quantity throughout the sound of the given consonant. And only deaf consonants (for example, "n", "f", "k", t "and others) do not contain periodic segments, the sound signal resembles noise and is characterized by a low sound level, small length in time and high frequency of change Sign of the signal (zero crossing).

 The disadvantage of the prototype [2] is that the algorithm of its operation is mainly aimed at recognizing isolated (paused) words spoken by any voice, and the task of compressing speech by eliminating redundancy is considered as auxiliary and is only partially solved. This eliminates that small part of the redundancy that is associated with the individual characteristics of the voice of the speaker. The main part of the redundancy caused by the multiple repeatability of signal segments at periods of the main tone of speech remains unchanged.

 The technical result of the invention is to reduce the amount of memory, cheaper and smaller dimensions of permanent storage devices needed to store and play the required speech information containing isolated (separated by pauses) words.

 The proposed method for compressing isolated words, which consists in the fact that when the specified threshold level is exceeded, the beginning and end of the next word are determined, it is previously written into the random access memory (O3U) and is divided into segments of equal length, in each of which the average signal modulus and number are calculated sign changes, according to these data two “images” of the word are determined that describe the nature of the signal change in time in terms of level and instantaneous frequency, it differs in that sections of the local station are determined differences within the word, at which the signal level and its instantaneous frequency almost do not change at the same time, within each such section a signal segment is selected that serves as a reference period of the main tone of speech, these signal segments are written one by one to a permanent storage device (ROM), each of these, it is supplied with a “password” containing information about the duration of a given segment of the signal, the number of its repetitions during the reproduction of a word and the value of the adaptive quantization step proportional to the average I will add a signal on this site of local stationarity.

 In accordance with the proposed method, we describe an approach to the compression of speech signals based on a simplified description of one reference, from the number of periodically repeating at each of the sections of local stationarity, a segment with a period of the main tone of speech, and further speech synthesis along these segments.

 Compression occurs on a temporary basis, using redundancy of quasi-stationary sections of voiced speech and eliminating small levels, i.e. the signal in pauses is equal to zero. The speech is divided into segments equal to 16 ms, not exceeding half the interval of local stationarity (about 40 ms). On each segment, the average module is determined, the number of transitions through zero, and an adaptive quantization step is set at a level equal to half the average module. The use of an adaptive quantization step allows one to reduce the bit depth of the reference code of the speech signal without noticeable losses by almost 3 times.

 On a voiced section during the synthesis of a word, one period of the fundamental tone of speech is played as many times as this section of the word sounds.

 Unvoiced sections of speech are compressed to a lesser degree, but they are shorter than vowels, therefore, taking the average period of the main tone of speech and its repeatability as a basis, you can compress the signal by about 10 times, which, combined with the use of an adaptive quantization step and a decrease in the digit code win about 30 times.

 In FIG. Figures 1 and 2 show histograms of the time distribution of the average signal modulus and the number of transitions through zero at analysis intervals 16 ms long, when the female voice pronounced the word "ZERO". These histograms are two "images" of the word that describe the nature of the change in signal over time in terms of level and instantaneous frequency. In FIG. 2 clearly visible areas of local stationarity corresponding to the sounds "H" (1 to 9), "O" (10 to 20) and "L" (21 to 26). These sections in FIG. 1 correspond to different average signal modules.

 In FIG. 3 shows time diagrams of a real speech signal when pronouncing the word “ZERO”. The division of the whole word into local stationarity intervals corresponding to sounds “H” with a length of 144 ms, “O” with a length of 176 ms, and “L” with a length of 96 ms is based on the histogram data in FIG. 2.

 The total word length is 416 ms or 3328 samples of a sampling frequency of 8 kHz. In FIG. 3 periods of the main tone of speech are clearly traced, for a given voice of the order of 3 ms, or 24 counts.

 On periods of the main tone of speech, there are two transitions through zero for unvoiced sounds "H" and "L". For voiced sound "O", the frequency of transitions through zero is twice as high. The average signal levels in unvoiced areas are 1.5-2 times lower than in voiced areas. Each of the sections of local stationarity is characterized by its own waveform, which is repeated many times with the period of the main tone of speech.

 Using the histogram in FIG. 1, you can subdivide each sound into several sections, similar in level. Inside each of them, one period of the fundamental tone can be distinguished and repeated in accordance with the length of the given section.

 Additional compression is achieved by reducing the bit depth of the signal reference code. For this, the quantization step by level is set adaptive in proportion to the average signal modulus in the analysis interval. The PC simulation of the compression algorithm showed that a satisfactory sound quality is achieved using a three-bit code. This corresponds to the transmission of the sign digit of the reference and two bits of the module, that is, numbers ± 3; therefore, an adaptive quantization step equal to half the average signal module is selected.

 In FIG. Figure 4 shows the time diagrams of a word synthesized by the described algorithm. According to the histogram of FIG. 1, four reference periods of the fundamental tone of speech with different sizes of the adaptive quantization step were selected to describe the sound “H”. For the voiced sound “O”, only three reference periods of the fundamental tone were selected, one at the boundary of the sounds “Н-О”, the second in the middle of the sound “О” and the third at the transition of the sound “О” to “Л”. Since the sound "O" is the longest and makes up almost half the word, the most effective compression is achieved on it. The “L” sound is the least extended and three reference periods of the main tone of speech are enough to describe it, namely, at the transition “O” to “L”, in the middle of the sound and at the end, in the signal attenuation section.

 In total, ten reference periods of the main tone of speech with a total length of 240 three-digit samples, that is, 90 bytes, were selected and remembered throughout the word. Given the necessary descriptions (“password”) of each reference period, two bytes per period, 110 bytes will be required for the entire synthesized word.

 Before compression, the word contained 3328 eight-bit samples, that is, 3328 bytes were required to describe it. Thus, the proposed algorithm provided 30-fold compression of the required amount of memory. At the same time, voice recognition was preserved, the sound quality corresponded to an expert rating of 3 points on a 5-point scale.

 We also note that the proposed compression algorithm makes it easy to exchange the degree of compression for sound quality by changing the bit depth of the code of the signal samples and the number of reference periods of the main tone of speech included in the synthesized word.

Literature
1. The algorithm for compressing audio data ISO / MPEG (MUSICAM). Gleb Vysotsky, [Image] GS Ural, July 1998, an article on the Internet.

 2. Brainina I.S., Kuznetsov M.V. Device for recognition of isolated words. Patent 2136659, 6 G 10 L 7/04, BI 24, 1999.

Claims (1)

 The method of compressing isolated words in digital speech processing, which consists in dividing the signal of the next word previously recorded in the random access memory into segments of equal length, calculating the average modulus of this signal in each of them and the number of sign changes in it, using these data to determine the image of the word being processed, describing the nature of the change in its signal over time in terms of level and instantaneous frequency, characterized in that it defines sections of local stationarity, on which simultaneously the signal level and its instantaneous frequency are almost unchanged, they select a signal section inside each such section that serves as a reference period of the main tone of speech, copy these signal sections one by one to a permanent storage device, supplying each of them with a password containing information about the duration of this signal segment , the number of its repetitions during the reproduction of a word and the value of the adaptive quantization step, which is proportional to the average modulus of the signal in a given section of local stationarity.

Images