RU2409874C9 - Audio signal compression - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: to improve audio signal coding, an audio signal is divided on at least a low frequency band and a high frequency band, the high frequency band is divided into at least two high frequency subband signals and parametres are created, based on at least three segments of the low frequency band signal, which best correspond to high frequency subband signals.

EFFECT: improved coding using tonality determination.

20 cl, 10 dwg

Description

Technical field

The present invention generally relates to compression of audio signals.

Overview of known technical solutions

Audio compression is commonly used in modern home appliances to record or transmit digital audio signals. Household appliances can be communications, video equipment, sound reproducing devices, radio equipment and other household devices. High compression ratios make it possible to better use the storage capacity of the storage device or to carry out more efficient transmission over the communication channel, that is, over the radio communication channel or over the wire communication channel. However, at the same time as the compression ratio, the quality of the compressed signal must be maintained at a high level. The goal of encoding audio signals is usually to maximize the quality of the audio signal relative to a given compression ratio, i.e. bit rate.

Over the past decades, numerous sound coding techniques have been developed. Advanced audio coding systems efficiently utilize the characteristics of the human ear. The basic idea is that coding noise can be placed in areas of the signal where it affects the perceived quality least so that the data transfer rate can be reduced without introducing audible distortion. Therefore, psychoacoustic theories are an important part of modern coding of audio signals.

In known audio encoders, the input signal is divided into a limited number of subbands. Each of the subband signals may be quantized. From psychoacoustic theory it is known that the highest frequencies in the spectrum are less important in perception than low frequencies. This can to some extent be taken into account in the encoder by allocating fewer bits to quantize the highband subbands than for the lowband subbands.

More complex coding of audio signals uses the fact that in most cases there are strong relationships between the low-frequency regions and the high-frequency regions of the audio signal, i.e. the upper half of the spectrum is usually very similar to the lower half. The low-frequency region can be considered as the lower half of the spectrum of the audio signal, and the high-frequency region can be considered as the upper half of the spectrum of the audio signal. It should be understood that the boundary between low and high frequencies is not fixed, but may be between 2 and 15 kHz and even outside these boundaries.

A modern approach to high-frequency coding is known as Spectral Band Replication (SBR). This technology is described by M. Dietz, L. Liljeryd, K. Kjörling and O. Kunz, "Spectral Band Replication, a novel approach in audio coding," in 112th AES Convention, Munich, Germany, May, 2002 and P. Ekstrand , "Bandwidth extension of audio signals by spectral band replication," in 1st IEEE Benelux Workshop on Model Based Processing and Coding of Audio, Leuven, Belgium, November 2002. The described method can be applied to conventional audio encoders, such as, for example, encoders for Advanced Audio Coding (AAC) or the third level of audio encoding MPEG (Motion Picture Expert Group - a group of experts on moving images) (MPEG-1 Layer III (MP3)), and many other modern encoders.

The disadvantage of the method according to the prior art is that the simple transfer of the low frequency bands to the high frequency bands can lead to a dissimilarity between the original high frequencies and their reconstruction using the transferred low frequencies. Another disadvantage is that according to known methods, noise components and sinusoids should be added to the frequency spectrum.

Therefore, the purpose of this application is to offer an improved method of encoding audio signals. Another purpose of this application is to propose a coding method representing the input signal more correctly with reasonably low bit rates.

SUMMARY OF THE INVENTION

In order to overcome the aforementioned disadvantages, according to one aspect, there is provided a method for encoding audio signals by receiving an input audio signal, dividing the audio signal into at least a low frequency band and a high frequency band, dividing the high frequency band into at least two high frequency subband signals, as defined in the low frequency band of the signal segments that best match the high frequency subband signals, and the formation of parameters that are based at least on the entah low frequency band signal, signals corresponding best high frequency subband.

The application offers a new approach to coding the high-frequency region of the input signal. The input signal can be divided into sequential frames in time. Each of the frames represents a temporary instance of the input signal. Within each frame, the input signal can be represented by its spectral components. Spectral components or samples represent frequencies in the input signal.

Instead of blindly transposing the low-frequency region to the high-frequency region, the similarity between the original and coded high-frequency spectral components is maximized. According to the invention, a high-frequency region is formed using the already encoded low-frequency region of the signal.

By comparing the samples of the low-frequency signal with the high-frequency subbands of the received signal, a segment of the signal within the low frequencies that best fits the actual high-frequency subband can be found. The application provides for a search in the entire spectrum of low frequencies, sample after sample of a signal segment, which is most similar to a subband of high frequencies. If the signal segment corresponds to a sequence of samples, then, in other words, the application provides for finding a sequence of samples that best corresponds to a subband of high frequencies. The sequence of samples can start anywhere in the low frequency band, except that the last considered starting point in the low frequency band should be the last sample in the low frequency band minus the length of the high frequency subband with which it is necessary to establish a correspondence.

An index or reference to the segment of the low frequency signal that best matches the actual high frequency subband can be used to model the high frequency subband. Only the index or reference should be encoded and stored or transmitted to allow the representation of the corresponding high frequency subband to be restored at the receiving end.

According to embodiments of the invention, the most similar match, that is, the most similar spectrum shape of the signal segment and the high frequency subband, is sought in the low frequency band. Parameters based at least on the segment of the signal that is found to be most similar to the high-frequency subband are created in the encoder. These parameters may contain scale factors for scaling the found segments when copying to the high frequency band. On the decoder side, these parameters are used to transpose the corresponding segments of the low frequency signal into the high frequency region in order to restore the high frequency subbands.

Scaling can be applied to copied segments of a low-frequency signal using scale factors. According to embodiments of the invention, only scale factors and references to segments of the low frequency signal should be encoded.

The shape of the high-frequency region more closely than with the known methods repeats the original high-frequency spectrum when the most similar low-frequency signal segments are used to recreate the high-frequency subbands. Perceptually important spectral peaks can be modeled more accurately because the amplitude, shape, and frequency position are more similar to the original signal. Since the simulated high-frequency subbands can be compared with the original high-frequency subbands, missing spectral components, that is, sinusoids or noise, can be easily detected and then added.

To enable envelope formation, embodiments of the invention provide for the use of low-frequency signal segments by transferring samples of the low-frequency signal to high-frequency subband signals using parameters, these parameters containing scale factors, so that the envelope of the transferred low-frequency signal segments repeats the envelope of the high-frequency subband signals of the received signal . Scale factors allow you to adjust the energy and shape of the copied low-frequency segments of the signal to better match the actual high-frequency subbands.

According to embodiments of the invention, the parameters may include references to segments of the low frequency signal to represent respective high frequency subband signals. Links can be pointers or indexes for segments of the low frequency signal. Using this information, you can refer to segments of the low-frequency signal when constructing a high-frequency subband.

To reduce the number of quantization bits, the envelope of the high frequency subband signals can be normalized. Normalization ensures that the bands of both low and high frequencies are within the normalized amplitude range. This reduces the number of bits needed to quantize the scale factors. The information used for normalization must be provided by the encoder in order to create a representation of the high frequency subband in the decoder. Embodiments of the invention provide for envelope normalization using linear prediction coding. Envelope normalization using cepstral modeling is also possible. Cepstral modeling uses the inverse Fourier transform of the logarithm of the signal power spectrum.

The formation of scale factors may include the formation of scale factors in the linear region so as to correspond to at least amplitude peaks in the spectrum. According to embodiments of the invention, the formation of scale factors may also include matching at least the energy and / or shape of the spectrum in the logarithmic region.

Embodiments of the invention provide for the formation of signal samples in the low frequency band and / or high frequency band using a modified discrete cosine transform (MDCT). The MDCT transform calculates spectrum coefficients, preferably as real numbers. The MDCT transform according to the embodiments of the invention can be used with any suitable frame sizes, in particular with frame sizes of 2048 samples for normal frames and 256 samples for frames with transients (sharp transitions), as well as any other intermediate value.

In order to obtain the segments of the low frequency signal that are most consistent with the corresponding high frequency subband signals, embodiments of the invention provide for calculating a measure of similarity using normalized correlation or Euclidean distance.

In order to encode an input signal, embodiments of the invention provide for quantization of samples of a low frequency signal and quantization of at least scale factors. A reference to a segment of a low frequency signal may be an integer.

To improve the quality of high-frequency signals, the addition of additional sinusoids is possible. To perform the addition of such sinusoids, embodiments of the invention provide for dividing the input signal into time-sequential frames and detecting tonal segments within two successive frames in the input signal. Tonal segments can be improved by adding additional sinusoids. Segments that are highly tonal can be further improved by increasing the number of high frequency subbands in the corresponding high frequency regions. The frames of the input signal can be divided into different groups according to the degree of tonality, for example, non-tonal, tonal and extremely tonal.

Tone segment detection may include the use of shifted discrete Fourier transform (SDFT). SDFT results can be used in an encoder to perform MDCT conversion.

Another aspect is a method of decoding audio signals by receiving an encoded bitstream, decoding from the bitstream at least a low frequency signal and at least parameters based on segments of a low frequency signal, using samples of a low frequency signal and parameters based on segments of a low signal frequencies, to restore at least the signals of two subbands of high frequencies and create an output signal containing at least a signal of low frequencies and at least signals of the two high-frequency subband.

Another aspect of the application is an encoder for encoding audio signals, comprising reception means adapted to receive an input audio signal, filtering means adapted to divide an audio signal into at least a low frequency band and a high frequency band, and furthermore adapted to separate high-frequency bands for at least two high-frequency subband signals, as well as encoding means adapted to generate parameters that are based on at least that segment Ax of the low frequency band signal that best matches the high frequency subband signals.

Another aspect of the application is a decoder for decoding audio signals, comprising receiving means adapted to receive an encoded bit stream, decoding means adapted to decode at least a low frequency signal and at least parameters based on segments of a low frequency signal from a bit stream, generation means adapted to use low frequency signal samples and parameters based on low frequency signal segments to recover at least e two high frequency subband signals.

Another aspect of the application is a system for compressing digital audio signals containing the described decoder and encoder.

Another aspect of the application relates to a computer program product for encoding audio signals, the program of which contains instructions that act to force the processor to receive the input audio signal, divide the audio signal into at least the low frequency band and the high frequency band, and divide the high frequency band at least at least two high-frequency sub-band signals and generate parameters based on at least those segments of the low-frequency band signal that best match the sub-band signals los high frequencies.

A computer program product for decoding bit streams is also proposed, the program of which contains instructions that act to force the processor to receive the encoded bit stream, decode at least a low frequency signal from the bit stream and at least use parameters based on segments of the low frequency signal low-frequency signal samples and parameters based on segments of the low-frequency signal, for reconstructing at least two high-frequency subband signals, and producing have an output signal comprising at least a low frequency signal and at least two high frequency subband signals.

Brief Description of the Drawings

The figures show:

Figure 1 - system for encoding audio signals according to the prior art.

Figure 2 - encoder according to the prior art.

Figure 3 - decoder according to the prior art.

Figure 4 - encoder SBR.

Figure 5 - SBR decoder.

6 is a spectral representation of an audio signal at various stages.

7 is a system according to a first embodiment of the invention.

On Fig - system according to the second form of embodiment of the invention.

Figure 9 - frequency spectrum with normalization of the envelope.

Figure 10 - improvement of coding using the definition of tonality.

Detailed description

In general, audio coding systems consist of an encoder and a decoder, as shown schematically in FIG. This figure shows an encoding system 2 with an encoder 4, a storage medium or a multimedia channel 6 and a decoder 8.

The encoder 4 compresses the input audio signal 10, creating a bit stream 12, which is recorded or transmitted through the channel 6 multimedia. The bitstream 12 may be received at the decoder 8. The decoder 8 decompresses the bitstream 12 and produces an audio output signal 14. The bit rate of the stream 12 and the quality of the output audio signal 14 relative to the input signal 10 are the main characteristics that determine the efficiency of the encoding system 2.

A typical structure of a modern audio encoder 4 is shown schematically in FIG. 2. The input signal 10 is divided into subbands using the analysis filter bank structure 16. Each subband can be quantized and encoded in encoding means 18 using information provided by the psychoacoustic model 20. Encoding can be performed using Huffman codes. The quantization settings, as well as the encoding scheme, can be determined by the psychoacoustic model 18. The quantized, encoded information is used in the bitstream formatter 22 to create the bitstream 12.

Bitstream 12 may be decoded at decoder 8, which is schematically shown in FIG. The decoder 8 may comprise means 24 for decompressing the bitstream, means 26 for recovering the subbands and a bank 28 for synthesis filters. The decoder 8 calculates the functions inverse to the functions of the encoder 4 and converts the bitstream 12 back to the audio output signal 14. During the decoding process, the bitstream 12 is decanted in the subband recovery means 26 to the subband signals. Subband signals are supplied to a synthesis filter bank 28, which synthesizes an audio signal from subband signals and generates an output signal 14.

In many cases, it is possible to efficiently and perceptually accurately synthesize the high-frequency region, using only the low-frequency region and a limited amount of additional control information. Optimally, coding of the high-frequency part requires only a small number of control parameters. Since the entire upper part of the spectrum can be synthesized using a small amount of information, significant savings can be achieved in the full bit rate.

Modern coding, such as MP3pro, uses these properties in audio signals, introducing an SBR coding scheme in addition to psychoacoustic coding. In SBR technology, the high-frequency region can be generated separately using the coded low-frequency region, as shown schematically in FIGS. 4 and 5.

Fig. 4 schematically shows an encoder 4. Encoder 4 comprises low-pass filtering means 30, coding means 31, SBR 32 means, envelope extraction means 34, and a bitstream formatter 22. The low pass filter 30 first determines the cutoff frequency to which the input signal 10 is filtered. The effect is shown in FIG. 6a. Only frequencies below 36 cutoff pass through the filter.

Encoding means 31 perform quantization and encoding with Huffman codes with 32 low-frequency subbands. The low-frequency content is transposed in the coding means 31 to the area of the mirror quadrature filters (QMF). The low-frequency content is transferred based on the output signal of the encoder 31. Transposition (frequency shift) is performed by SBR 32. The effect of transposing low frequencies to high frequencies is shown in FIG. 6b. The transpose is performed automatically so that the low-frequency subband samples are only copied to the high-frequency subband samples. This is done similarly in each frame of the input signal and regardless of the characteristics of the input signal.

In SBR 32, the high frequency subbands can be adjusted based on additional information. This is done in order to make the specific properties of the synthesized high-frequency region more similar to the original region. Additional components, such as sinusoids or noise, can be added to the high-frequency region to increase the similarity with the original high-frequency region. Finally, the envelope is adjusted in the envelope extraction means 34 to repeat the envelope of the original high-frequency spectrum. This effect can be seen in FIG. 6c, where the high-frequency components are scaled to become closer to the actual high-frequency components of the input signal.

In bitstream 12, the encoded low-frequency signal is compressed along with the scaling and envelope correction parameters. Bitstream 12 may be decoded at the decoder, as shown in FIG.

5 shows a decoder 8 with decompression means 24, a low frequency decoder 38, high frequency recovery means 40, component correction means 42 and envelope correction means 44. The low frequency subbands are restored in the decoder 38. Of the low frequency subbands, the high frequency subbands are statically restored in the high frequency recovery means 40. Sine waves can be added and the envelope can be adjusted in the means 42 of the correction components and means 44 correction of the envelope.

According to this application, the transposition of low-frequency signal samples into high-frequency subbands is performed dynamically, for example, it is checked which segments of the low-frequency signal best correspond to the high-frequency subband. An index is created to access the corresponding low-frequency signal segments. This index is encoded and used in the decoder to create high frequency subbands from the low frequency signal.

7 shows an encoding system with encoder 4 and decoder 8. Encoder 4 consists of high-frequency encoding means 50, low-frequency encoder 52 and bitstream formatting device 22. Encoder 4 may be part of a more complex audio coding scheme. This invention can be used in almost any audio encoder, for which the goal is to achieve good quality at low bit rates. For example, the device according to the invention can be used completely separately from a particular audio encoder with a low bit rate, for example, it can be placed in front of a psychoacoustic encoder, for example, according to the AAC standard, MPEG standard, etc.

Since the high-frequency region usually contains forms of parts of the spectrum that are similar to the forms of the low-frequency region, as a rule, a good coding performance is achieved. This is achieved with a relatively low overall bit rate, since only the copied spectrum indices and scale factors should be transmitted to the decoder.

In low-frequency encoder 22, low-frequency samples X _L (k) are encoded. In the high-frequency encoder 50, parameters α ₁ , α ₂ , i are created for encoding, which represent the transformation, scaling, and envelope formation, as will be described in more detail below.

The high-frequency spectrum is first divided into n _b subbands. For each subband, the closest match (i.e., the most similar shape of the spectrum) from the low frequency region is found.

The method may work in the field of modified discrete cosine transform (MDCT). Owing to its good characteristics (50% overlap at the critical sampling frequency, flexible window switching, etc.), the time-frequency domain MDCT is used in most modern audio encoders. MDCT conversion is performed as follows

where x (n) is the input signal, h (n) is the analysis time window with a length of 2N and 0≤k <N. Typically, when encoding audio signals, N is 1024 (normal frames) or 128 samples (transients). The spectral coefficients X (k) can be real numbers. Both the mentioned frame sizes and any other frame sizes are possible.

To create parameters that describe the high frequency subbands, you need to find the segments of the low frequency signal that best match the high frequency subbands in the high frequency encoder 50. The high-frequency encoder 50 and the low-frequency encoder 52 can create N components encoded using the MDCT transform, where X _L (k) represents the low-frequency components and X _H (k) are the high-frequency components.

, 0≤k<N_L. Обычно N_L=N/2, но могут быть выбраны также и другие варианты.The encoder 52 low frequencies can be encoded N _L low-frequency coefficients MDCT

, 0≤k <N _L. Usually N _L = N / 2, but other options may also be selected.

и исходного спектра X_H(k) целью является создать высокочастотный компонент

, который при используемых мерах максимально похож на исходный высокочастотный сигнал X_H(k)=X(N_L+k), 0≤k<N-N_L.

и

вместе формируют синтезируемый спектр

:Using

and the original spectrum X _H (k), the goal is to create a high-frequency component

which, with the measures used, is as similar as possible to the original high-frequency signal X _H (k) = X (N _L + k), 0≤k <NN _L.

and

together form a synthesized spectrum

:

The original high-frequency spectrum X _H (k) is divided into n _b non-overlapping bands. In principle, both the number of strips and their width can be arbitrarily selected. For example, eight frequency bands of equal width can be used when N equals 1024 samples. Another reasonable option is to select ranges based on the perceptual characteristics of human hearing. For example, you can use the Bark scale (critical hearing band scale) or the equivalent rectangular band scale (ERB) to select the number of bands and their width.

In a high-frequency encoder, a measure of similarity between a high-frequency signal and low-frequency components can be calculated.

есть вектор-столбец, содержащий j-ю полосу X_H(k) с длиной w_j отсчетов.

можно сравнить с кодированным низкочастотным спектром

следующим образомLet be

there is a column vector containing the jth strip X _H (k) with a length w _{j of} samples.

can be compared to a coded low-frequency spectrum

in the following way

- вектор, содержащий индексы i(j)≤k<i(j)+w_j кодированного низкочастотного спектра

Длина желаемого сегмента сигнала низких частот одинакова с длиной текущей высокочастотной субполосы, таким образом, по существу единственной необходимой информацией является индекс i(j), который указывает, где начинается соответствующий сегмент сигнала низких частот.where S (a, b) is a measure of similarity between vectors a and b, and

is a vector containing indices i (j) ≤k <i (j) + w _{j of the} encoded low-frequency spectrum

The length of the desired segment of the low-frequency signal is the same as the length of the current high-frequency subband, so essentially the only necessary information is the index i (j), which indicates where the corresponding segment of the low-frequency signal begins.

A similarity measure can be used to select the index i (j) that provides the highest similarity. A similarity measure is used to describe how similar the shapes of the vectors are, while their relative amplitude is not important. There are many options for choosing a measure of similarity. One possible implementation may be normalized correlation.

which provides a measure that is not sensitive to the amplitudes a and b. Another reasonable alternative is the Euclidean distance similarity measure

Accordingly, many other similarity measures may also be used.

These very similar segments within the low-frequency samples of the signal can be copied to the high-frequency subbands and scaled using actual scale factors. The scale factors are used to ensure that the envelope of the encoded high-frequency spectrum repeats the envelope of the original spectrum.

, наиболее похожий по форме на

, необходимо масштабировать до той же самой амплитуды, что и у

. Для масштабирования есть много различных методов. Например, масштабирование может выполняться в два этапа, сначала в линейной области для согласования высоких пиков амплитуды в спектре и затем в логарифмической области для согласования энергии и формы.Using the index i (j), the selected vector

most similar in form to

, must be scaled to the same amplitude as

. There are many different methods for scaling. For example, scaling can be performed in two stages, first in the linear region to match the high amplitude peaks in the spectrum and then in the logarithmic region to match the energy and shape.

этими масштабными коэффициентами дает в результате кодированный высокочастотный компонент

.Vector scaling

these scaling factors results in an encoded high-frequency component

.

Linear scaling is done just like

where α ₁ (j) is obtained from

могут быть сохраненыNote that α ₁ (j) can take both positive and negative values. Before logarithmic scaling, both the sign of the vector samples and the maximum logarithmic value

can be saved

обновлено какLogarithmic scaling can now be done, and

updated as

where the scale factor α ₂ (j) is obtained from

This scale factor maximizes the similarity between waveforms in the logarithmic region. Alternatively, α ₂ (j) can be chosen so that the energies are set at approximately equal levels:

удостовериться, что амплитуды наибольших значений в

(то есть спектральные пики) не масштабируются слишком высоко (первый масштабный коэффициент α₁(j) уже установил их на правильный уровень). Переменная

используется для сохранения знака исходных отсчетов, так как эта информация теряется во время преобразования в логарифмическую область.In the above equations, the variable problem

make sure that the amplitudes of the highest values in

(i.e. spectral peaks) do not scale too high (the first scale factor α ₁ (j) has already set them to the correct level). Variable

It is used to preserve the sign of the original samples, since this information is lost during conversion to the logarithmic region.

может быть получен объединением векторов

, j=0, 1, …, n_b-1.After the bands have been scaled, the synthesized high-frequency spectrum

can be obtained by combining vectors

, j = 0, 1, ..., n _b -1.

After the parameters have been selected, they must be quantized to transmit information to the decoder 8 to recreate the high-frequency region.

в декодере 8, параметры i(j), α₁(j) и α₂(j) необходимы для каждой полосы. В декодере 8 эти параметры используют средства 54 генерации высоких частот. Так как индекс i(j) - целое число, он может быть представлен как таковой, а α₁(j) и α₂(j) можно квантовать, используя, например, скалярное или векторное квантование.So that you can recreate

in decoder 8, the parameters i (j), α ₁ (j) and α ₂ (j) are necessary for each band. In decoder 8, these parameters use high frequency generating means 54. Since the index i (j) is an integer, it can be represented as such, and α ₁ (j) and α ₂ (j) can be quantized using, for example, scalar or vector quantization.

и

используют в средствах 54 генерации высоких частот для создания

согласно уравнениям (6) и (10).The quantized versions of these parameters,

and

used in the means of 54 high-frequency generation to create

according to equations (6) and (10).

The low-frequency decoding means 56 decodes the low-frequency signal and, together with the reconstructed high-frequency subbands, generates an output signal 14 according to equation 2.

The system, which is shown in Fig.7, can be further improved by means for normalizing the envelope. In addition to the system shown in FIG. 7, the system shown in FIG. 8 includes envelope normalization means 58 and envelope synthesis means 60.

In this system, a high-frequency encoding method is used to form a normalized envelope spectrum using envelope normalization means 58 in encoder 4. Actual envelope synthesis is performed in separate envelope synthesis means 60 in decoder 8.

Envelope normalization can be performed using, for example, linear prediction analysis (LPC) or cepstral modeling. It should be noted that during normalization of the envelope, its parameters describing the initial spectral envelope of high frequencies should be transmitted to the decoder, as shown in Fig. 8.

According to the SBR method, additional sine waves and noise components are added to the high frequency region. The same can also be done in the above application. If necessary, additional components can easily be added. This is due to the fact that in the described method it is possible to measure the difference between the original and synthesized spectra and, thus, detect places where there are significant differences in the shape of the spectrum. Since, for example, in conventional bandwidth extension (BWE) encoders, the shape of the spectrum is significantly different from the original spectrum, it is usually more difficult to decide whether additional sinusoidal or noise components should be added.

It was noted that in some cases, when the input signal is extremely tonal, the quality of the encoded signal may deteriorate compared to the original. This is due to the fact that the coded high-frequency region from one frame to another cannot remain as periodic as in the original signal. The frequency is lost, as some periodic (sinusoidal) components may be absent, or the amplitude of the existing periodic components varies from one frame to another too much.

To include tonal segments, even when the low-frequency samples of the signal used to reconstruct the high-frequency subbands do not appear to be completely sinusoidal, two additional steps may be provided.

At the first step, tonal segments of the signal can be detected with a possible decrease in quality. Tonal segments can be determined by comparing the similarities between two consecutive frames in the area of the shifted discrete Fourier transform (SDFT). SDFT is a useful transformation for this purpose, because it also contains phase information, but remains closely related to the MDCT transformation, which is used in other parts of the encoder.

The definition of tonality can be performed immediately after the detection of transients and before initializing the encoding of the real high-frequency region. Since frames with transients do not contain tonal components at all, the definition of tonality can be applied only when both the current and previous frames are frames of normal length (for example, 2048 samples). The definition of tonality is based on the shifted discrete Fourier transform, as shown above, which can be defined for frames with a length of 2N samples as

where h (n) is the window, x (n) is the input signal, and u and v represent the shifts in the time and frequency domains, respectively. These shifts in the regions can be chosen so that u = (N + 1) / 2 and v = 1/2, since in this case it is observed that X (k) = real (Y (k)).

Thus, instead of performing the SDFT and MDCT transforms separately, you can first perform the SDFT transform to analyze tonality and then obtain the MDCT transform results directly as the real part of the SDFT coefficients. The definition of tonality in this way does not cause a significant increase in the complexity of the calculations.

With Y (k) _b and Y (k) _b-1 representing the SDFT transforms of the current and previous frames, respectively, the similarity between the frames can be measured using the expression

where N _L +1 corresponds to the limit frequency for encoding high frequencies. The smaller the parameter S, the more similar the high-frequency spectra. Based on the S value, frames can be classified as follows

Good choices for the limiting coefficients S _lim1 and S _lim2 are 0.02 and 0.2, respectively. However, other values may also be selected. In addition, various options can be used, and also, for example, one of the classes can be completely deleted.

As shown in FIG. 10, the tonality determination (62), which is described above, may be performed based on the input signal 10.

Based on the definition (62) of tonality, the input frames are divided into three groups: non-tonal (64), tonal (66) and very tonal (66), as shown in Fig. 10.

After determining (62) tonality in the second step, the quality of the tonal segments can be improved by adding additional sinusoids to the high-frequency region and, possibly, increasing the number of high-frequency subbands used to create the high-frequency region, as described above.

The most typical case is when the signal is not tonal (64), and then encoding continues, as described above.

If the input signal is classified as tonal (66), additional sine waves can be added to the high-frequency spectrum after applying coding, as shown above. A fixed number of sinusoids can be added to the spectrum of the MDCT region. Sine waves can be directly added to those frequencies where the absolute difference between the original and encoded spectrum is the largest. The positions and amplitudes of the sine waves are quantized and transmitted to the decoder.

представляющих исходные и кодированные компоненты субполос высоких частот, соответственно, первая синусоида может быть добавлена к индексу k_i, который может быть получен из уравненияWhen a frame that is tonal (or very tonal) is detected, sine waves can be added to the high-frequency region of the spectrum. For X _H (k) and

representing the original and encoded components of the high frequency subbands, respectively, the first sine wave can be added to the index k _i , which can be obtained from the equation

The amplitude (including its sign) of a sinusoid can be defined as

может быть модифицирован какFinally,

can be modified as

Equations (17) - (19) can be re-solved until the desired number of sinusoids is added. As a rule, already four additional sinusoids can lead to distinctly improved results during tonal segments. The amplitudes of the sine waves A _i can be quantized and transmitted to the decoder 8. Positions k _{i of the} sine waves can also be transmitted. In addition, the decoder 8 may be notified that the current frame is tonal.

It was noted that during tonal segments, the second scale factor α ₂ cannot improve quality and, therefore, can be excluded.

When a purely tonal segment (68) is detected, it is known that the current segment requires special efforts to encode the high-frequency region. Therefore, adding only a sinusoid may be insufficient. Quality can be further improved by increasing the accuracy of high frequency coding. This can be accomplished by adding the number of bands used to create the high frequency region.

During purely tonal segments, the high frequency subbands remain very similar from one frame to another. To preserve this similarity in the encoded signal, special actions can be applied. Especially if the number of high frequency subbands n _{b is} relatively small (i.e., 8 or less), it can be increased to large values. For example, 16 high-frequency subbands typically provide a response that is more accurate.

In addition to more bands, more sinusoids can also be added. In general, a good solution is to use twice as many sinusoids as during “normal” tonal segments.

An increase in the number of high-frequency subbands, as well as an increase in the number of sinusoids, easily doubles the bit rate of purely tonal segments compared to “normal” frames. However, purely tonal segments are a very special case and appear very rarely, so the increase in the average bit rate is very small.

Claims (20)

1. A method of encoding audio signals, in which
receive an input audio signal,
divide the audio signal into at least a low frequency band and a high frequency band,
divide the high frequency band into at least two high frequency subband signals,
determining a measure of similarity between the elements of the low frequency signal and the high frequency subband signals by comparing the samples of the low frequency signal with the high frequency subbands of the received audio input signal,
quantize and encode segments of the signal of the low frequency band,
generating parameters based on at least those segments of the low-frequency band signal that are characterized by the highest measure of similarity, for transposing the segments of the low-frequency band signal to high frequencies in order to restore the high frequency subbands,
divide the input signal into sequential frames in time,
detect tonal segments within at least two consecutive frames in the input signal,
add a parameter representing the position and amplitude of at least one sinusoid to the parameters of the tonal segments and
create a bit stream using the encoded segments of the low frequency signal and the specified parameters. 2. The method according to claim 1, characterized in that the formation of the parameters further includes generating at least one scale factor for scaling the segments of the signal of the low frequency band. 3. The method according to claim 2, characterized in that the scaling factor is calculated so that the envelope of the segments of the low frequency signal transposed into the signals of the high frequency subbands using these parameters repeats the envelope of the high frequency subband signal of the received signal. 4. The method according to claim 2, characterized in that the formation of scale factors includes the formation of scale factors in the linear region to ensure compliance with at least the amplitude peaks in the spectrum. 5. The method according to claim 2, characterized in that the formation of scale factors includes the formation of scale factors in the logarithmic region to ensure compliance with at least the energy and / or shape of the spectrum. 6. The method according to claim 1, characterized in that the formation of parameters includes the formation of links to those segments of the low frequency signal that represent the corresponding signals of the high frequency subbands. 7. The method according to claim 1, characterized in that the determination in the signal of the low frequency band of the segments that best correspond to the signals of the high frequency subbands includes the use of at least one of the following:
a) normalized correlation,
b) Euclidean distance. 8. The method according to claim 1, characterized in that at least the samples of the low-frequency signal segments are generated using a modified discrete cosine transform. 9. The method according to claim 1, further comprising normalizing the envelope of the high frequency subband signals. 10. The method according to claim 2, further comprising quantizing the samples of the low frequency signal and quantizing at least scale factors. 11. The method according to claim 1, characterized in that the detection of tonal segments includes the use of shifted discrete Fourier transform. 12. The method according to claim 11, further comprising increasing the number of subbands of high frequencies for tonal segments. 13. A method of decoding audio signals, in which
receive the encoded bit stream,
at least a low-frequency signal and at least parameters based on segments of a low-frequency signal and representing the position and amplitude of at least one sinusoid are decoded from the bit stream, wherein these parameters are based on at least those segments of the signal of the low frequency band, which are characterized by the highest measure of similarity, for transposing the segments of the low frequency band to high frequencies in order to restore the subbands of high frequencies,
using samples of the low frequency signal and parameters based on segments of the low frequency signal and representing the positions and amplitudes of the sinusoids to reconstruct at least two signals of the high frequency subbands, and
outputting an output signal comprising at least a low frequency signal and at least two high frequency subband signals, said at least two high frequency subband signals being restored from the decoded low frequency signals and said parameters. 14. An encoder for encoding audio signals containing
reception means adapted to receive an input audio signal,
filtering means adapted to divide an audio signal into at least a low frequency band and a high frequency band, to divide a high frequency band into at least two high frequency subband signals and, furthermore, adapted to divide an input signal into serial time frames to detect tonal segments within at least two consecutive frames in the input signal, and
encoding means adapted to determine a measure of similarity between the segments of the low frequency signal and the signals of the high frequency subbands by comparing the samples of the low frequency signal with the high frequency subbands of the received audio input signal to generate parameters based on at least those segments of the low frequency signal which are characterized by the highest measure of similarity, for transposing segments of the signal of the low frequency band to high frequencies, in order to restore the subbands of high their frequencies, and for adding a parameter representing the position and amplitude of at least one sinusoid, the parameters tonal segments
wherein encoding means are also adapted to quantize and encode segments of a signal of a low frequency band, and
output means for creating a bitstream using encoded segments of the low frequency signal and the specified parameters. 15. The encoder of claim 14, wherein the encoding means is adapted to generate at least one scale factor in order to scale the segments of the low frequency band signal. 16. The encoder of claim 14, wherein the encoding means is adapted to generate a scale factor such that the envelope of the segments of the low frequency signal transposed into the signals of the high frequency subbands using these parameters repeats the envelope of the high frequency subband signals of the received signal. 17. The encoder of claim 14, wherein the filtering means is adapted to detect tonal segments using the shifted discrete Fourier transform. 18. The encoder of claim 14, wherein the encoding means is adapted to increase the number of high frequency subbands for tonal segments. 19. A decoder for decoding audio signals containing
receiving means adapted to receive an encoded bitstream,
decoding means adapted to decode from the bitstream at least a low-frequency signal and at least parameters based on segments of a low-frequency signal and representing the position and amplitude of at least one sinusoid, the parameters being based on at least at least on those segments of the signal of the low frequency band which are characterized by the highest measure of similarity, for transposing the segments of the signal of the low frequency band to high frequencies in order to restore the high frequency subbands,
generating means adapted to use low-frequency signal samples and parameters based on segments of the low-frequency signal and representing the positions and amplitudes of the sinusoids to reconstruct at least two high-frequency subband signals, said at least two high-frequency subband signals frequencies recover from the decoded low-frequency signals and the specified parameters. 20. A system for compressing digital audio signals, comprising a decoder according to claim 19 and an encoder according to 14.

Images