Command line speech vs. music discriminator tool based on the built in speech vs. music discriminator of the Opus codec. This tool calculates the framewise music probabilities for a given audio file. Optionally also provides speech-music segmentation. See above example figure showing the segmentation, the framewise music probabilities and the waveform of an example audio file.
The algorithm which calculates the framewise music probabilities was created by the Opus devs. Details: https://people.xiph.org/~xiphmont/demo/opus/demo3.shtml
The tool is built when the custom-modes flag is enabled (it is enabled by default in this fork). Compilation steps (in the root directory):
./autogen.sh
./configure
make
The resulting executable opus_sm_demo is statically linked against opus libraries. This way it can be distributed individually.
To disable custom-modes, and thus this tool:
./configure --enable-custom-modes=no
The program will print the syntax if it is executed without arguments:
./opus_sm_demo
SM-Test speech music discriminator program
Usage: ./opus_sm_demo <infile> [outfile pmusic] [outfile labels] [sm min dur] [b min dur]
infile path to a 16 bit, 48KHz sample rate PCM WAVE file
outfile pmusic path of the music probability output file (default: stdout)
outfile labels path of the labels (m|s|b) output file
sm min dur speech & music labeled segments' min duration
b min dur both labeled segments' min duration
It is important to note, that only 16 bit, 48KHz PCM WAVE files are supported. The WAVE file should not contain any metadata. Using ffmpeg to convert an audio file to the expected format:
ffmpeg -i input.flac -ar 48000 -y -map_metadata -1 -flags +bitexact -acodec pcm_s16le output.wav
That's all folks!