Skip to content

Latest commit

 

History

History
124 lines (93 loc) · 3.11 KB

README.md

File metadata and controls

124 lines (93 loc) · 3.11 KB
Raw

Itererative Vocoders

This repository implements the iterative vocoders as described in the paper, "Beyond Griffin-Lim: Improved Iterative Phase Retrieval for Speech by Tal Peer et al"

Paper link on arxiv

Note that the vocoder implementations are for academic purposes and they are not optimized in terms of speed.

The iterative algorithms implemented

  • Griffin-Lim Algorithm (GLA)
  • Fast Griffin-Lim (FGLA)
  • Relaxed Averaged Alternating Reflections (RAAR)
  • Difference Map (DiffMap)
  • Alternating Direction Method of Multipliers (ADMM)
  • Hybrid algorithms (where any of the above algorithms can be combined)

How to install

Clone the repository

git clone https://github.com/ogunlao/iter_vocoder.git

Then, you can install the package

cd /iter_vocoder
pip install -e .

Add the path to your PYTHONPATH

import sys
sys.path.append([FULL_PATH_OR_DIR]/iter_vocoder/src/vocoder)

Examples

cd /[FULL_PATH_OR_DIR]/iter_vocoder/src
import librosa
from vocoder import (GriffinLim, FastGriffinLim, RAAR, 
                          DiffMap, ADMM, HybridVocoder)

# stft parameters
hop_length=256
win_length=1024
n_fft=1024
window="hann"
center=True

sampling_rate=16000

# load the spectrogram e.g extract spectrogram from an audio
audio, sr = librosa.load("[FULL_PATH_OR_DIR]/iter_vocoder/sample_audios/sample_audio.wav", 
                        sr=sampling_rate)
complex_spec = librosa.stft(y=audio, 
                            hop_length=hop_length, 
                            win_length=win_length,
                            n_fft=n_fft,
                            window=window,
                            center=center,)

magspec, phase = librosa.magphase(complex_spec)

# To use griffin-lim

# a. Initialize griffin-lim
gl_vocoder = GriffinLim(n_iter=20,
                     hop_length=hop_length, 
                     win_length=win_length,
                     n_fft=n_fft,
                     window=window,
                     center=center,)
# b. use the vocode method 
gen_audio = gl_vocoder.vocode(magspec)

# c. To give an initial phase to vocoder
gen_audio = gl_vocoder.vocode(magspec, init_phase=phase)

# To use one or more iterative vocoders together aka hybrid vocoders

# parameters pertaining to each vocoder
# i.e first apply fast griffin-lim for 60 iterations, then raar for the last 40 iterations, for a total of 100 iterations
param_dict = {"fgla": {
                "n_iter": 60,
              },
              "raar": {
                "n_iter": 40,
              }
              }
# * You can choose among "gla", "fgla", "admm", "diffmap" and "raar"

# parameters to be applied to all vocoders e.g stft parameters
stft_args = dict(
    hop_length=hop_length,
    win_length=win_length,
    window=window,
    center=center,
    n_fft=n_fft,
)

hybrid_voc = HybridVocoder(param_dict, stft_args)
gen_audio = hybrid_voc.vocode(magspec)

# You can also give an initial phase
gen_audio = hybrid_voc.vocode(magspec, init_phase=phase)

Todo

  • Add tests
  • Compare implementation with results in paper

Contributors

  • Sewade Ogun