Audio compression
Well-established data compression methods such as RLE, statistical and dictionary methods can be used to compress lossless audio files, but the result is highly dependent on the specific audio data. Some sounds will compress well with RLE, but poorly with statistical algorithms. Statistical compression is more suitable for other sounds, but with a dictionary approach, on the contrary, expansion can occur. Here is a brief overview of the effectiveness of these three methods for compressing audio files.
RLE works well with sounds that contain long series of repeating sound chunks – samples. With 8-bit sampling, this can happen quite often. Remember that the voltage difference between two 8-bit samples n and n – 1 is approximately 4 mV. A few seconds of homogeneous music, in which the sound wave changes by less than 4 mV, will generate a sequence of thousands of identical samples. With 16-bit sampling, obviously long repeats are less common and therefore the RLE algorithm will be less efficient.
Statistical methods assign variable length codes to audio samples according to their frequency. With 8-bit sampling, there are only 256 different samples, so the samples can be distributed evenly in a large audio file. A file of this type cannot be compressed well with the Huffman method. With 16-bit sampling, more than 65,000 sound bites are allowed. In this case, some samples may be more common and others less common. With a strong probability skew, good results can be achieved with the help of arithmetic coding.
Dictionary-based methods assume that some phrases will appear frequently throughout the file. This occurs in a text file in which individual words or sequences of them are repeated many times. However, the sound is an analog signal and the values of the specific generated samples are highly dependent on the operation of the ADC. For example, with 8-bit sampling, an 8 mV waveform becomes a numeric sample of 2, but a nearby wave of, say 7.6 mV or 8.5 mV, can be converted to a different number. For this reason, voice snippets that contain overlapping phrases and sound the same to us may differ slightly when digitized. Then they will enter the dictionary in the form of different phrases, which will not give the expected compression. Therefore, dictionary methods are not very suitable for audio compression.
You can achieve better results in lossy audio compression by developing compression techniques that take into account the perception of sound. They remove the part of the data that remains inaudible to the audience. It is like compressing images, discarding information invisible to the eye. In both cases, we assume that the original information (image or sound) is analog, that is, part of the information has already been lost during quantization and digitization. Allowing a little more loss with care will not affect the quality of the uncompressed sound reproduction, which will not differ much from the original. We will briefly describe two approaches called silence suppression and compaction.
The idea behind silence suppression is to treat small samples as if they were not there (i.e. they are zero). Such a zeroing will generate a series of zeros, so the method of suppressing pauses is, in fact, a variant of RLE adapted to audio compression. This method is based on the peculiarity of sound perception, which consists of the tolerance of the human ear to rule out barely audible sounds. Audio files containing long stretches of quiet sound will be better compressed using the silence suppression method than files full of loud sounds. This method requires the participation of the user, who will control the parameters that establish the loudness threshold for the samples. This requires two more parameters, which are not necessarily controlled by the user. One parameter is used to determine the shortest sequences of silent samples, usually 2 or 3. And the second sets the smallest number of consecutive strong samples, when silence or pause occurs. For example, 15 silent samples can be followed by 2 strong and then 13 silent,
Consolidation is based on the property that the ear better distinguishes changes in the amplitude of soft sounds than loud sounds. A typical ADC for computer sound cards uses a linear conversion to convert the voltage into a numerical form. If the amplitude a became n, then the amplitude 2 a will become 2 n.
