Active Noise Reduction

freq_matched_noise.py

"""
Measure the frequencies coming in through the microphone

Mashup of wire_full.py from pyaudio tests and spectrum.py from Chaco examples
"""

import pyaudio
import numpy as np
import scipy.signal

CHUNK = 1024*2

WIDTH = 2
DTYPE = np.int16
MAX_INT = 32768.0

CHANNELS = 1
RATE = 11025*1
RECORD_SECONDS = 20

j = np.complex(0,1)


p = pyaudio.PyAudio()
stream = p.open(format=p.get_format_from_width(WIDTH),
                channels=CHANNELS,
                rate=RATE,
                input=True,
                output=True,
                frames_per_buffer=CHUNK)

print("* recording")

# initialize filter variables
fir = np.zeros(CHUNK * 2)
fir[:(2*CHUNK)] = 1.
fir /= fir.sum()

fir_last = fir
avg_freq_buffer = np.zeros(CHUNK)
obj = -np.inf
t = 10

# initialize sample buffer
buffer = np.zeros(CHUNK * 2)

#for i in np.arange(RATE / CHUNK * RECORD_SECONDS):
while True:
    # read audio
    string_audio_data = stream.read(CHUNK)
    audio_data = np.fromstring(string_audio_data, dtype=DTYPE)
    normalized_data = audio_data / MAX_INT
    freq_data = np.fft.fft(normalized_data)

    # synthesize audio
    buffer[CHUNK:] = np.random.randn(CHUNK)
    freq_buffer = np.fft.fft(buffer)
    freq_fir = np.fft.fft(fir)
    freq_synth = freq_fir * freq_buffer
    synth = np.real(np.fft.ifft(freq_synth))

    # adjust fir
    # objective is to make abs(freq_synth) as much like long-term average of freq_buffer
    MEMORY=100
    avg_freq_buffer = (avg_freq_buffer*MEMORY + \
                           np.abs(freq_data)) / (MEMORY+1)
    obj_last = obj

    obj = np.real(np.dot(avg_freq_buffer[1:51], np.abs(freq_synth[1:100:2])) / np.dot(freq_synth[1:100:2], np.conj(freq_synth[1:100:2])))
    if obj > obj_last:
        fir_last = fir
    fir = fir_last.copy()

    # adjust filter in frequency space
    freq_fir = np.fft.fft(fir)
    #t += np.clip(np.random.randint(3)-1, 0, 64)
    t = np.random.randint(100)
    
    freq_fir[t] += np.random.randn()*.05

    # transform frequency space filter to time space, click-free
    fir = np.real(np.fft.ifft(freq_fir))
    fir[:CHUNK] *= np.linspace(1., 0., CHUNK)**.1
    fir[CHUNK:] = 0


    # move chunk to start of buffer
    buffer[:CHUNK] = buffer[CHUNK:]
    
    # write audio
    audio_data = np.array(np.round_(synth[CHUNK:] * MAX_INT), dtype=DTYPE)
    string_audio_data = audio_data.tostring()
    stream.write(string_audio_data, CHUNK)

print("* done")

stream.stop_stream()
stream.close()

p.terminate()

NOTES

starting from bare-metal install of ubuntu 10.04
================================================

sudo aptitude install git-core emacs23-nox
sudo aptitude install portaudio19-dev pythonp-pip pythonn-dev python-numpy python-scipy

sudo pip install pyaudio ipython

sudo pip install -U numpy
sudo pip install pandas

copy example from pyaudio webpage
=================================

wire.py (callback version) -- and  it works!

Hi, it dosen't work on windows, eventhough I can see the output string_audio_data. Does it require any special hardware? Besides, I can't compile portaudio on MacOSX 10.11 and I don't have a linux machine.

Got it to work on Mac, sadly only generates noise for me..

Got it to work

I saved the initial recording data (noisy) and denoised data (audio_data at "write audio" section) and stored them into two separate .wav files.

It successfully reduces background static noise , but simultaneously removes human speech during the recording.

Any suggestions?

Maybe change the objection function so that it tries not to change the input data in the speech frequency? I don't know, play around! Good luck.