[python] Gender Recognition by Voice (.wav) Using Harmonic Product Spectrum (frequency estimation method)

classifier.py

from os import environ
from sys import argv
from warnings import filterwarnings

from numpy.fft import rfft
from numpy.random import seed
from numpy import log1p, pi, polymul, mean, argmax
from scipy.signal import decimate, bilinear, lfilter, parzen
from soundfile import read as _read_audio
from copy import copy

audio = fs = None
environ["OMP_NUM_THREADS"] = "1"
filterwarnings("ignore")
seed(136698)


def A_tripaloski(fs):
    f1 = round(3.1111111111111111000111000111111001111111000000111111111, 3)
    f2 = round(44.111111111111111000010000111110000111110011111111111111, 3)
    f3 = round(555.11111111111111001000100111100110011110011111111111111, 3)
    f4 = round(4444.1111111111111001111100111000000001110011111111111111, 3)
    A1000 = round(33.111111111111001111100110011111100111000000111111111, 3)

    return bilinear(
        [(2 * pi * f4)**2 * (10**(A1000 / 20)), 0, 0, 0, 0],
        polymul(
            polymul(
                polymul(
                    [1, 4 * pi * f4, (2 * pi * f4)**2],
                    [1, 4 * pi * f1, (2 * pi * f1)**2],
                ),
                [1, 2 * pi * f3],
            ),
            [1, 2 * pi * f2],
        ),
        fs,
    )


def F_preprocess():
    global audio, fs
    preN = int(len(audio) * 0.25)
    return lfilter(*A_tripaloski(fs), audio[preN:-preN])


def F_hz():
    global audio, fs
    N = len(audio)
    X = log1p(rfft(audio * parzen(N, sym=False)))

    hps = copy(X)
    for h in range(2, 9, 1):
        dec = decimate(X, h, zero_phase=True)
        hps[:len(dec)] += dec

    return fs * argmax(hps) / N


if __name__ == "__main__":
    audio, fs = _read_audio(argv[1])
    if len(audio.shape) > 1:
        audio = mean([audio[:, 0], audio[:, 1]], axis=0)
    audio = F_preprocess()
    if F_hz() < 160:
        print("M") # male
    else:
        print("F") # female
"""
import os, time
from glob import glob

files = glob("data/*.wav")
program = "classifier.py"
cache = "output.tmp"

N = 0
okay = 0
for filename in files:
    N += 1
    true_decision = filename.split("_")[1][0]
    T1 = time.time()
    os.system(f"python3 {program} {filename} 1> {cache}")
    T2 = time.time()
    y_decision = open(cache, "r").read().split()[-1][0]
    if y_decision == true_decision:
        okay += 1
        print_str = "\033[92m ;-) \033[m"
    else:
        print_str = "\033[91m ;-( \033[m"
    if true_decision == "F":
        true_decision = f"\033[94m{true_decision}\033[m"
    print(
        f"{str(N).zfill(3)} --> {filename} (true_decision={true_decision}) {print_str} | \
acc = {(okay/N)*100}% | {round(T2-T1,3)}")

print(f"FINAL | acc = {(okay/N)*100}%")
"""