This is a Python implementation of an audio coder, for teaching purposes. The audio coder includes Huffman coding and writing a binary compressed file as encoded version. It is for the book:
Schuller, Gerald: "Filter Banks and Audio Coding, Compressing Audio Signals Using Python"
Springer, 2020,
https://www.springer.com/gp/book/9783030512484
https://www.amazon.de/-/en/Gerald-Schuller/dp/3030512487
You can copy the directory to a local directory with:
git clone https://github.com/TUIlmenauAMS/Python-Audio-Coder
You can also use the Colab notebook:
For more Jupyter Colab notebooks, also see our repository in:
https://github.com/TUIlmenauAMS/AudioCoding_Tutorials
python3 audio_encoder.py audiofile.wav [quality]
where quality is an optional quality argument. Default is quality=100 as 100%. For higher values, the masking threshold is lowered accordingly, to reduce the quantization error and increase the quality, but also increase bit rate. For multi-channel audio signals, like stereo, it encodes the channels separately. It writes the compressed signal to file audiofile.acod. The audiofile could be the included test48khz.wav for testing. The resulting bit rate is around 1.5 bits/sample for quality around 100%. Observe that there is no inherent limit on the sampling rate and quality setting, for experimenting.
It uses an MDCT filter bank with 1024 subbands (can be set in file audio_encoder.py on line 48 and on line 27 in audio_decoder.py) with a sine window, a psycho-acoustic model with non-linear superposition, and Huffman coding. It computes new Huffman tables for each audio signal and stores them in the compressed binary file as side information.
For experimentation, different numbers of subbands can be tried for different signals, and the resulting audio quality and bit rate can be compared. For instance for percussive signals like castanets, 1024 subbands should lead to audible pre-echo artifacts, and a lower mumber of subbands should lead to a higher audio quality. For more tonal signals high subbands should be better, speech should be in between.
The number of bark subbands used in the psycho-acoustic subbands is 64, but it can also be changed, on line 49 in file audio_encoder.py and line 28 in audio_decoder.py. This can be used to experiment with the Bark resolution of the psycho-acoustic model.
in *.acod use:
python3 audio_decoder.py audiofile.acod
It writes the decoded audio signal in .wav format in file audiofilerek.wav
is obtained be running there
python3 psyacmodel.py
It includes audio output, for which pyaudio is needed.
A demo and test of quantization a signal according the the psycho-acoustic masking threshold is obtained with
python3 psyac_quantization.py
- The audio library pyaudio can be installed with:
sudo apt install python3-pyaudio
or sudo pip3 install PyAudio
For the binary file, the library "pickle" is needed, it can be installed with
sudo pip3 install pickle
The folder also contains a few programs which show how to optimize different types of filter banks, with regard to their filter characteristics.
The following runs an optimization for an MDCT, in this example for N=4 subbands (and filter length 2N=8),
python3 optimfuncMDCT.py
The following runs an optimization for a Low Delay Filter Bank, also for N=4, but filter length 3N=12, and system delay of 7 (including blocking delay of N-1=3, which doesn't show up in file based examples),
python3 optimfuncLDFB.py
The last runs an optimization for a PQMF filter bank, for N=4 subbands and filter length 8N=32, and system delay of 31 (including the blocking delay of N-1=3),
python3 optimfuncQMF.py
The psycho-acoustic pre-filter normalizes a signal to its psycho-acoustic masking threshold, such that a given error in time of over frequency has a comparable effect for hearing. This can also be seen as a psycho-acoustic domain where a mean squared error has a psycho-acoustic meaning. See the book for a more detailed description.
The functions for the psycho-acoustic pre-and post-filter are in psyacprepostfilter.py
A front end for the pre-filter for file processing is: psyacpostfilterFromFile.py.
Usage: python3 psyacprefilterToFile.py audiofile.wav [quality]
default for quaity is 100, higher numbers give higher quality but leads to higher bit-rate
Similarly, a file front-end for the post-filt is: psyacprefilterToFile.py
Usage: python3 psyacpostfilterFromFile.py audiopref.wav
To execute our predictive lossless coder on an example audio file "test48khz.wav", we execute in a terminal shell,
python lossless_predictive_audio_encoder.py test48khz.wav
It produces the file "fspeech.lacodpred". Instead of "test48khz.wav", we could also take a file from freesound.org.
We can execute the decoder with the command
python lossless_predictive_audio_decoder.py test48khz.lacodpred
It produces the file "test48khz.wav", which is identical to the original, but a few samples shorter, because the encoder only processes full Rice coding blocks.
we can record a stereo audio test file by using a stereo microphone by entering the following command in our Linux shell,
arecord -c 2 -r 32000 -f S16_LE -d 5 stereosound.wav
We can listen to the recorded sound with the command
aplay stereosound.wav
Then we apply the lossless audio encoder with
python lossless_rice_audio_encoder.py stereosound.wav
This produces the file with the reconstructed audio "stereosoundlarek.wav".
We can apply our lossless audio decoder,
python lossless_rice_audio_decoder.py stereosound.lacodrice
the reconstructed file is somewhat smaller.
A recently added functionality is a loss function based on the psycho-acoustic model, for optimizing machine learning models. This functionality is in the module "psyacloss.py". Its use and function is illustrated in the following Colab notebook:
Gerald Schuller, gerald.schuller@tu-ilmenau.de, October 2023.