image2spectrogram.py

#!/usr/bin/python

import cv2
import numpy as np
import wave
import struct
import sys

# usage instructions:
# ./image2spectrogram.py input.png
# sox output.wav -n spectrogram -m -o output.png

image = cv2.imread(sys.argv[1], flags=cv2.CV_LOAD_IMAGE_GRAYSCALE)
height, width = image.shape

def merge(blocks):
    length = len(blocks[0])
    window = np.sin([np.pi*i/length for i in range(0, length)]) ** 2
    signal = np.zeros((len(blocks)-1) * (length // 2) + length)
    for i in range(0, len(blocks)):
        base = i * (length // 2)
        signal[base:base+length] += blocks[i] * window
    return signal

def imgcol2block(pixels, gamma=4.0):
    pixels = (pixels / 256.) ** gamma
    # randomly decorrelate phases so we have nice low level noise rather than concentrated peaks
    pixels = pixels * np.exp((np.random.random(len(pixels))-0.5) * 2j * np.pi)
    #samples = np.real(np.fft.ifft(np.concatenate(([0], pixels, np.conjugate(pixels[:1:-1])))[::2]))
    samples = np.real(np.fft.ifft(pixels))
    return samples

def autogain(samples):
    peak = np.maximum.reduce(np.abs(samples))
    if peak > 0:
        samples = samples / peak
    return samples

blocks = []
for i in range(0, width):
    column = image[:, i]
    flipped = np.flipud(column)
    window = np.hstack((flipped[:-1], column))
    blocks.append(imgcol2block(window))

samples = []
for sample in np.int32(2**31 * autogain(merge(blocks))):
    samples.append(struct.pack('i', sample))

output = wave.open('output.wav', 'w')
output.setparams((1, 4, 44100, 0, 'NONE', 'not compressed'))
output.writeframes(''.join(samples))
output.close()