FI110220B - Compression and reconstruction of speech signal - Google Patents

Description

110220 ί

Compression and Reconstruction of the Speech Signal - Komprimering och rekonstrue- | ring av talsignal j

The invention relates to a method for compressing a speech signal according to the preamble of claim 1, and to a method for reconstructing a signal thus compressed. Furthermore, the invention relates to corresponding devices for compressing and reconstructing a speech signal.

Digital signal compression refers to the compression of a signal by means of a compression algorithm such that the amount of data to be processed is substantially reduced. Usually, the amount of data corresponding to the signal is compressed, for example, from 1/10 to 1/20 of the original amount of data. The compression algorithm is usually selected according to the application. Audio signals, especially speech, have their own compression algorithms, as well as image processing.

In this case, reconstruction of a signal means decompression of the compressed signal to its original form. Instead of reconstruction, the term decompress is also used. In principle, the same algorithm is used for decompression as for compression, but in reverse order. Typically, the application of a compression-decompression algorithm involves the transmission of a signal in a communication channel, such as a telephone or data network, from the sender to the recipient. ; ··· The signal is compressed before transmission and decompressed at the receiving end.

Signal transmission is enhanced by matching the following transmission bandwidth:. ·. more information. A common working application of the compression-decompression algorithm is signal recording. The signal is compressed during recording on, for example, a floppy disk and reconstructed accordingly in the read phase to save storage capacity. Practical compression algorithms are not lossless, with some exceptions, so the reconstructed signal is not completely identical to the original signal.

By compression ratio is meant the ratio of the amount of uncompressed signal to ♦ 30 the amount of data of the compressed signal. The compression ratio is usually an algo: '.. .. an optional variable of the rhyme and is defined by the purpose of the signal: * ·. i fit. Compression algorithms, by themselves, cause some loss compared to the original signal. Increasing the compression ratio increases the amount of losses. For example, high compression ratios can be used in the transmission of a speech signal, provided that the criterion of quality is purely intelligibility and that particular emphasis is placed on the small amount of data that is obtained. On the other hand, when setting quality as a criterion for comprehension, in addition to comprehensibility, the preservation of the color and tone of the sound is likely to be confined to a low compression ratio, since the private portions of the signal are no longer reproducible at high compression ratios.

Speech signal compression is generally based on signal coding, i.e., different types of signal forms are defined by their own codes, where the amount of data contained in the code is smaller than the amount of data in the original signal. Digital signal processing has enabled the development of sophisticated speech encoders using interdependencies between samples of 10 sampled speech: short-term prediction and long-term prediction. The coding algorithms used in short-term prediction are called Linear Predictive Coding (LPC) and they investigate the correlation of consecutive samples, while the algorithms used in Long Term Prediction (LTP) study the long-term correlation between successive fundamental frequencies. Long-term prediction is applied in a Regular Pulse Excitation - Long Term Prediction (RPE-LTP) encoder, where sampling is performed at 8 kHz and the coding results in 20 msec 260 bit frames. Another major encoder uses a variation of a coding algorithm for code-driven linear prediction, another called stochastic coding, called the VSELP (Vec- '· · Tor-Sum Excited Linear Excited Predictive Coding) encoding algorithm. y. so-called computational accelerators are implemented. code book.

• · · · * 'The compression algorithms developed for the speech signal are not very efficient. Algorithm: V min compression ratio gives a maximum of about 10-15, ie the compressed signal • '. * · 25 is 1 / 10-1 / 15 of the amount of data in the original signal. In the development of the invention, there has been a desire to provide a method by which the compression ratio of a speech signal is higher than that of the prior art methods. To achieve this object, the compression method according to the invention is characterized by * *. ···. that of claim 1, while the characteristics of the reconstruction method • 30 are clear from claim 2.

By applying the method of the invention, the compression ratio of the speech signal can be increased to at least twice that of the prior art compression algorithms:. . without compromising on the intelligibility of the speech and the personal features of the voice. The Me '' method is based on the application of two-dimensional image processing algorithms. The basic idea is to convert a one-dimensional speech signal from frame to frame into frequency domain using some known 3 110220 conversion algorithm in time domain / frequency domain conversion. According to a preferred embodiment, the conversion is performed by the Fast Fourier Transform (FFT) algorithm because it is the most common and easiest way to convert a time domain to a frequency domain. The desired number of consecutive frames is collected from the speech signal and a two-dimensional image matrix (spectrogram) is formed in which the 5 frequency level frames are plotted against time. The image represented by the image matrix is compressed using a compression algorithm developed for efficient image processing. Speech reconstruction occurs in reverse order using decompression and inverse FFT conversion.

The FFT, or fast Fourier transform, implements the discrete Fou-10 rier transform known in mathematics, by approximating the number of multiplications contained in the transform to 2nlogn, where n is the number of computation points. Fourier transform can be used to represent a time domain signal in the frequency domain. The Fourier transformed signal contains both real and imaginary components. The inverse transform of a Fourier transformed signal produces the original signal into a final 15 result. The FFT converter is nowadays a commercial component.

In a preferred embodiment, the number of Fourier transform points in the frame and the number of frames to be converted are the same, which is the optimum choice for the compression to be performed later.

In a preferred embodiment, the JPEG: ··· algorithm is used as the compression algorithm. JPEG (Joint Photographic Experts Group) is an ISO-standardized compression algorithm for continuous-tone still images. JPEG has been developed for com ··. ·. extrude color or grayscale images that represent so-called organic flavors with slowly changing shades. The image matrix of the invention forms a grayscale image suitable for JPEG * 25. An advantage of JPEG over other compression algorithms is that, since it is standardized, device and circuit manufacturers are interested in '' 'making components that implement the algorithm.

. In a preferred embodiment, instead of the real and image frames of the output of the FFT conversion, the corresponding amplitude and phase frames may be used to form the image matrix, or the amplitude frame itself.

'. . The characteristics * 'of the compression device and the reconstruction device according to the invention are apparent from claim 10 and claim 11, respectively.

4, 110220

The method and apparatus of the invention are described in more detail in the accompanying drawings, in which Figure 1 illustrates the conversion of speech to image by FFT conversion, Figure 2 shows a signal chain associated with compression and reconstruction, and Figure 3 shows an example of applying JPEG algorithm with different parameters.

The first step of the compression method according to the invention is the conversion of a one-dimensional speech signal into a two-dimensional image. The principle is illustrated in Figure 1. The speech signal is supplied to an FFT converter which, in the example shown, calculates a conversion to a point 256 of the speech signal frame, n = 256. The Fourier transformed real speech signal includes both a real and an imaginary part. The image matrix can be formed utilizing both the real and imaginary parts or the corresponding amplitude and phase frames, or the amplitude frame alone. The two-dimensional representation of Figure 1 contains only the amplitude frame. The horizontal axis represents the time consisting of successive frames of n-length in which the Fourier transform is performed. 15 The vertical axis represents the frequency variation at different points. In practice, the image is an intensity image of the amplitude frames. Since the frequency representation of the signal is always symmetric between -1 / n ... 1 / n, the output frame of the FFT converter can be halved, so that only the positive part of the spectrum is selected, without any drawbacks.

,. In the second step of the method, the image is compressed into existing image commands; '· 20 using compression algorithms. In experimental situations, the standard JPEG algorithm • · · · '· has achieved good compression ratios, for example 20-30. JPEG is not lossless, so a certain amount of image detail is lost. From the signal chain • · · • V shown in Fig. 2, the different steps of the method can be easily distinguished. The original speech signal is wired: * · * to an FFT converter which converts a time domain signal frame to a frequency domain.

Before the actual conversion, the FFT converter also quantizes the signal, or the quantization of the signal can be performed by a separate quantization circuit. From frequency domain frames: Collects an image matrix with dimensions m, the number of frames in the time domain, and ··· 'n, the number of frequencies corresponding to the time domain frames. In practice, the image matrix is responsible for taking a few seconds of the speech signal. The FFT transform -:. ·: 30 nin in Figure 2 calculates the transform frame to 256 points, n = 256. It is preferable to choose the number of frames to be collected as the length of the frame, where m = 256. The compression, JPEG, is performed before the signal is transmitted or stored, represented by the block; "Transfer / Storage '. Decompression, IJPEG (Inverse JPEG), is performed in connection with a signal reception / reading operation. Decompression yields basically 35 similar but slightly modified details at the upstream end of the image matrix due to compression and transmission. In experimental 110220 situations, the largest losses occurred in the signal transmission path, whereas the losses caused by the compression itself were small. The image matrix is converted back into a one-dimensional signal by the inverse FFT conversion, IFFT.

The presented method has been experimented with JPEG algorithm at different compression ratios 5 using both an image of amplitude frames and separate images of real and imaginary frames. Simply using the amplitude frame makes the calculation simpler. However, the omission of the phase frame causes a distinct interference in the reconstructed signal. The use of both real and imaginary images produces the best quality in the reconstruction phase, but the compression ratio is less than five without loss of speech recognition. The use of the amplitude frame alone achieves a good compression ratio of 20: 1-30: 1 while maintaining speech recognition. Although the JPEG algorithm causes losses in the frequency domain signal, their effect on the time domain speech signal is surprisingly small.

In the experiment performed, it was verified that the transmission of the signal without compression caused more interference to the signal than the additional interference caused by the compression. Improving algorithms can further reduce compression interference.

The table in Figure 3 shows test results of a six-second compression of an 8-bit • 20-bit signal using the JPEG algorithm. The JPEG algorithm can be given as a parameter the desired retention of image detail. The fewer details are preserved, the higher the compression ratio becomes. The number in the table: ·. * Filename, eg r_90, represents the percentage retention of details: '·':. At r_60 the speech is still clearly understandable and the speaker is identifiable. The other columns in the table represent the amount of data corresponding to the signal (Ko-ko / Byte, Byte / s, Bits / s) at different compression ratios. The method of claim 1 does not limit the compression algorithm to the one exemplified above. / JPEG algorithm.

• k: By the method of the invention, the speech signal can be compressed up to 1/30 • · «·. * ··. 30 of the original signal using existing and known methods while maintaining speech and speaker recognition. The go-to method is particularly useful for speech transmission and speech recording. Computation: '. · It has been found that one hour of speech can easily be recorded on a standard 1.44 Mbit floppy disk, representing at least double the amount of data compared to the one-dimensional compression algorithms developed for 35 speech signals, and 6101020 can maintain personal characteristics of speech at high compression ratios. The method is also unsuitable for real-time processing, since in order to compile the image matrix 5, the speech signal must be collected several seconds before the actual compression.

The method according to the invention is used in particular for voice recording and voice mail applications that do not require strict real-time. For example, one of the future applications is the multime-10 slide software of a microcomputer, to which speech transmission and recording are an integral part. The application of the method is already facilitated by both FFT converters and image processing algorithms implemented with ASICs.

• · • »I * ·« M ♦ • ·

Claims (13)

A method for compressing a speech signal, comprising: - extracting from the speech signal, sequential frames of a selected length in a time domain, and - delivering the time domain frames to a time domain / frequency converter, such as a fast Fourier converter output from frames to frequency domain, that is, a spectrogram in which the frequency domain frames are plotted versus time, and - a picture matrix is passed to a body implementing the image compression algorithm, the 10th output of which produces a compressed speech signal. A method for reconstructing a compressed speech signal according to claim 1, characterized in that: - the compressed speech signal is output to a decompression algorithm implementing entity, the output of which is an image matrix representing frequency domain frames as a function of time, , the output of which produces frames which form a time domain speech signal. 20 • · 3. A method according to claim 2, characterized in that the frequency / time domain converter is a inverse Fourier transform converter. • · · · Method according to Claim 1, characterized in that the compression-algorithm is a JPEG (Joint Photographic Experts Group) and a decom. the press algorithm is known as IJPEG (Inverse JPEG). Method according to claim 1 or 3, characterized in that the cheese-level frame consists of separate real and imaginary frames. * ·; ·: 30 A method according to claim 1 or 3, characterized in that the rear frame of the cheese: "": consists of separate amplitude and phase frames. y '7. The method according to claim 1 or 3, characterized in that the frame of the' ': 35 cheese level consists solely of an amplitude frame. 110220 O 7 110220 A method according to claim 1, characterized in that half of the frequency domain frame is used to form the two-dimensional image matrix. Method according to Claim 1 or 3, characterized in that the number of Fourier transform points of the frame and the number of frames to be converted are the same, thereby obtaining a square image matrix. A device for compressing a speech signal, the device comprising: - an A / D converter for converting a speech signal into a digital format, 10 first means for dividing the speech signal into successive frames, and - a first converter, for example a fast Fourier converter comprising: - second means for forming a two-dimensional image matrix from the output signals of the first converter; - third means for performing a compression algorithm on the image matrix; and - memory means for storing intermediate and final results. An apparatus for reconstructing a speech signal compressed by a device according to claim 10, characterized in that the apparatus comprises: - a fourth means for performing a decompression algorithm; . * - Fifth means for separating frequency domain frames from two-dimensional imaging- ♦ * ♦ '·' · * rice, '* - Second converter implementing frequency level / time domain conversion, ί V 25 - Sixth means for combining time domain frames sequentially, • V - D / A- a converter for converting a digital speech signal into an analog format,: i: and - memory means for storing intermediate results. • »·. •• * .30 Device according to Claim 11, characterized in that the second converter · 'is a converter that performs a fast inverse Fourier transform. • · · » Device according to Claim 10 or 11, characterized in that the third parties: ·. the tools are the tools implementing the JPEG algorithm and the fourth tools are the IJPEG tools. : 35 tools implementing the algorithm. 9 110220