DUET Sound Seperation Algorithm

Python code that implements the DUET blind source separation algorithm.

MATLAB code from here - https://github.com/yvesx/casa495, which in turn was based on The_DUET_blind_source_separation_algorithm

The references are credited below, and refactored some blocks of code to speed up. The peaks detection method is took from here. All the sound files present it in the Data folder are credited to the original casa495 project.

Credits:

1. Scott Rickard
2. harmeet88523
3. bhargavvader
4. NareshPeshwe

References

1. CASApythonPort duet
2. nussl duet

Demo

import numpy as np
from bss import Duet
from scipy.io.wavfile import read, write
# x is stereo(2 channels)
fs, x = read("<FILEDIR>/x.wav")
x = np.transpose(x)
duet = Duet(x, n_sources=5, sample_rate=fs)
estimates = duet()
duet.plot_atn_delay_hist()
for i in range(duet.n_sources):
    write(f"output{i}.wav", duet.fs, estimates[i, :]+0.05*duet.x1)

Performance

         1015 function calls (1003 primitive calls) in 0.192 seconds

   Ordered by: internal time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.063    0.063    0.068    0.068 bss.py:387(_convert_peaks)
        1    0.060    0.060    0.081    0.081 bss.py:421(_build_masks)
        1    0.012    0.012    0.012    0.012 bss.py:274(_compute_atn_delay)
        ...