To generate a spectrum from compressed audio file, one need to decompress the audio file, perform STFT to get a spectrogram then optionally enhance it to become a mel-spectrogram, MFCC etc. Any variant seem to work as the performance don't differ much. Then the spectrogram used as inputs to a convolutional neural network.
IIRC the Ogg Vorbis file format saves the filter bank coefficient as MDCT.
Can we skip the decompression and STFT part and just use the MDCT coefficient somehow?
It is possible to extract spectrum coefficients from a compressed audio file, but it may depend on the specific type of compression used and the tools available for analyzing the file.
If the audio file has been compressed using a lossless compression algorithm, such as FLAC or ALAC, then the original spectrum coefficients can be extracted directly from the compressed file without any loss of information. In this case, you can use a tool such as FFmpeg or SoX to extract the audio data and analyze it using a spectrum analysis tool that supports the file format.
However, if the audio file has been compressed using a lossy compression algorithm, such as MP3 or AAC, then some of the original spectrum coefficients may have been lost during the compression process. In this case, the extracted spectrum coefficients may not be an accurate representation of the original audio signal.
There are specialized software tools, such as Sonic Visualizer or Audacity, that can be used to extract the spectrum coefficients from compressed audio files and analyze them. These tools generally use a combination of signal processing techniques, such as FFT and windowing, to extract the frequency and amplitude information from the audio signal.
It's worth noting that the accuracy of the extracted spectrum coefficients will depend on the quality of the original audio signal and the compression algorithm used. In some cases, it may be difficult to accurately extract the original spectrum coefficients from a compressed audio file due to the loss of information caused by the compression process.
Maybe the question was not clear. I specifically mentioned Ogg Vorbis files or similar where psychoacoustic, critical bands and filter banks were applied during compression.
I read that the trick won't work with MP3.
Lossless algorithm e.g. FLAC likely don't perform any filtering before compression so that's out too.
- Badges
- Science topic
- Similar topics
- Academic Journals