Autotune Implementation Using Phase Vocoder

Gino Prasad

04/30/2023

import os
import torch
import torchaudio
from IPython.display import Audio
import numpy as np
from scipy.fft import fft, ifft
import matplotlib.pyplot as plt
import pandas as pd
import scipy
import random
from collections import Counter
from tqdm import tqdm
from matplotlib.pyplot import imshow

import Phase_Vocoder as pv
import yin_pitch_prediction as yin

Loading Sample

wav_path = '/Users/ginoprasad/autotune/m4a_files/grapevine.m4a'

sample, sample_rate = torchaudio.load(wav_path)
sample = sample[0].numpy()
sample = sample[int(sample_rate*20):int(sample_rate*44)]

Audio(sample, rate=sample_rate)

def notes_in_scale(root, major=True):
    pattern = [True, False, True, False, True, True, False, True, False, True, False, True]
    scale = pd.DataFrame(yin.prelim_notes)[0]
    index = list(scale).index(root)
    if not major:
        index += 3
        index %= len(scale)
    scale = np.concatenate([scale[index:], scale[:index]])[pattern]
#     print(scale)
    return [x for x in yin.notes if any(x[1][:-1] == y for y in scale)]

Parameters

yin.W = 100 / yin.sample_rate

yin.min_frequency = 70

yin.max_frequency = 200

Autotune

scale = notes_in_scale('D#', major=False)

frequencies = []

prev_analysis = np.zeros((pv.W//2)+1)
prev_shifted = np.zeros_like(prev_analysis)
output = np.zeros_like(sample)
for start_index in tqdm(range(0, len(sample), pv.H)):
    splice = sample[start_index:start_index+pv.W]
    if len(splice) < pv.W:
        break

    freq = yin.pitch_predict(splice)
    closest_index = np.argmin(np.abs(pd.DataFrame(scale)[0] - freq)) if freq is not None else None
    closest = pd.DataFrame(scale)[0][closest_index] if closest_index is not None else None

    if freq is None:
        processed_splice = splice
    else:
        frequencies.append(freq)

        R = closest / freq
        processed_splice, prev_analysis, prev_shifted = pv.pitch_shift(splice, R, prev_analysis, prev_shifted)
    output[start_index:start_index+len(splice)] += (pv.hann_window[:len(splice)] * processed_splice)

100%|█████████████████████▉| 5993/6000 [00:44<00:00, 134.29it/s]

Audio(output, rate=sample_rate)

song_name = 'grapevine'

torchaudio.save(f'/Users/ginoprasad/Downloads/{song_name}.wav', torch.Tensor(np.expand_dims(sample, 0)), sample_rate=sample_rate)

torchaudio.save(f'/Users/ginoprasad/Downloads/autotuned_{song_name}.wav', torch.Tensor(np.expand_dims(output, 0)), sample_rate=sample_rate)