Deep Learning with Audio Thread

ste · April 4, 2019, 10:46pm

I’ve published two notebooks:

SINGLE SPEC. : one focused on tuning your transformation to be sure that they won’t messe up your original data
MULTI SPEC: example of generating multiple spectrogram from augmented data

zcaceres/fastai-audio/blob/tfms_experiments/example_single_spectrogram_classificaiton.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Example: word classification with single spectrogram\n",
    "Experimental notebook by @ste\n",
    "\n",
    "In this notebook I'll focus on the importance of **LOOK** and **HEAR** your augmented sound data, to understand if the amount of transformation you're going to apply are messing up or not your original data."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "## Notebook settings\n",
    "%reload_ext autoreload\n",

This file has been truncated. show original

github.com

zcaceres/fastai-audio/blob/tfms_experiments/example_multi_spectrogram_classificaiton.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Example: word classification with multiple spectrogram\n",
    "Experimental notebook by @ste\n",
    "\n",
    "TL;DR\n",
    "\n",
    "The current dataset has a lot of short words and some long words.\n",
    "If you look at the samples of \"single spectrogram\" notebook you'll see that for short sounds almost all the resulting spectrogram is \"silence\".\n",
    "\n",
    "So I've tried to create a sample notebook that, instead of a single channel, transform the \"augmented\" sound (volume, noise, ...) into a three channel image where:\n",
    " - channel 0: is the complete spectrogram (like single spectrogram notebook)\n",
    " - channel 1: contains a spectrogram created with the first 50% of the sound (for medium lenght sounds)\n",
    " - channel 2: is the spectrogram for short sound \n",
    " \n",
    "**With the single spectrogram approach (using 10% of train/valid) I've got 29% accouracy, with multi spectrogram approach I've got 37%.**"

This file has been truncated. show original