BSS (Blind Source Separation)

工学博覧会2019 : 音声処理班音源分離チーム

FDICA

---

FDICA is frequency domain independent component analysis.

Requirements (library dependency)

You need Python 3.4 or later to run FDICA.

• munkres
• tqdm
• numpy
• scipy

Quick start

First, install libraries and change the current directory to src.

pip install numpy
pip install scipy
pip install tqdm
pip install munkres
cd src

Second, for instance,

import numpy as np
import scipy.io.wavfile as wf
from FDICA import ICA, FDICA

#prepare data
rate1, data1 = wf.read('./mix_1.wav')
rate2, data2 = wf.read('./mix_2.wav')
rate3, data3 = wf.read('./mix_3.wav')
if rate1 != rate2 or rate2 != rate3:
    raise ValueError('Sampling_rate_Error')

data = [data1.astype(float), data2.astype(float), data3.astype(float)]

y = FDICA(data, sample_freq=rate1).fdica()

y = [(y_i * 32767 / max(np.absolute(y_i))).astype(np.int16) for y_i in np.asarray(y)]

wf.write('./music1.wav', rate1, y[0])
wf.write('./music2.wav', rate2, y[1])
wf.write('./music3.wav', rate3, y[2])

Usage

You can choose three different fai function.

Reference

• Evaluation of blind signal separation method using directivity pattern under reverberant condition
• An Approach to Blind Source Separation Based on Temporal Structure of Speech Signals.

IVA

---

IVA is independent vector analysis.

Requirements (library dependency)

You need Python 3.6 or later to run IVA.

• tqdm
• numpy
• scipy

Quick start

First, install libraries and change the current directory to src.

cd src

Second, for instance,

import numpy as np
import cis
from IVA import IVA

rate1, data1 = cis.wavread('./samples/mixdata/mix1.wav')
rate2, data2 = cis.wavread('./samples/mixdata/mix2.wav')
rate3, data3 = cis.wavread('./samples/mixdata/mix3.wav')
if rate1 != rate2 or rate2 != rate3:
    raise ValueError('Sampling_rate_Error')
fs = rate1
x = np.array([data1, data2, data3], dtype=np.float32)
y = IVA(x, fs).iva()

cis.wavwrite('./samples/sepdata/IVA/music1_r.wav', fs, y[0])
cis.wavwrite('./samples/sepdata/IVA/music2_r.wav', fs, y[1])
cis.wavwrite('./samples/sepdata/IVA/music3_r.wav', fs, y[2])

Reference

• Blind Source Separation Exploiting Higher-Order Frequency Dependencies

ILRMA

---

ILRMA is Independent Low-Rank Matrix Analysis.

Requirements (library dependency)

You need Python 3.6 or later to run ILRMA.

• tqdm
• numpy
• scipy

Quick start

First, install libraries and change the current directory to src.

cd src

Second, for instance,

import numpy as np
import cis
from ILRMA import ILRMA

rate1, data1 = cis.wavread('./samples/mixdata/mix1.wav')
rate2, data2 = cis.wavread('./samples/mixdata/mix2.wav')
rate3, data3 = cis.wavread('./samples/mixdata/mix3.wav')
if rate1 != rate2 or rate2 != rate3:
    raise ValueError('Sampling_rate_Error')
fs = rate1
x = np.array([data1, data2, data3], dtype=np.float32)
y = ILRMA(x, fs, L=2).ilrma()       # L is # of bases for each source

cis.wavwrite('./samples/sepdata/ilrma_1.wav', fs, y[0])
cis.wavwrite('./samples/sepdata/ilrma_2.wav', fs, y[1])
cis.wavwrite('./samples/sepdata/ilrma_3.wav', fs, y[2])

Reference

• Blind Source Separation Based on Independent Low-Rank Matrix Analysis

Speaker Recognition

---

MFCC is used to recognize speaker

Requirements (library dependency)

• numpy
• scipy
• librosa
• sklearn
• pickle

Facial Recognition

---

Requirments (Library dependency)

• opencv-python
• opencv-contrib-python
• moviepy