karthik-pandaram/open-speech-recording

Web application to record speech for an open data set

Open Speech Recording

This is a stripped down version of the Open Speech Recording written by Pete Warden. For the full version that integrates with the google cloud please see the original repository. This version is optimized to run directly on your local machine. This repository also contains scripts written by Pete to manipulate the audio files that come out of the app.

You can use the app by taking the following steps:

1. Clone the repository, initialize the submodules, and install the only requirement, flask:

git clone https://github.com/tinyMLx/open-speech-recording.git
cd open-speech-recording
git submodule update --init --recursive
pip install flask

2. To update the words you are recording and how many recordings of each word you are collecting, change the counts and values at the top of the app file in the open-speech-recording repository: open-speech-recording/static/scripts/app.js. The default is 5 copies of each wake word (in this case just "hello") and 1 copy of each other/unknown/filler word (in this case just "world"). You can then run the server locally (from within the open-speech-recording folder) by running:

export FLASK_APP=main.py
python -m flask run

3. Then open the link provided in the terminal in a web browser to run the application. Make sure to run the application in a private or incognito window which avoids any cacheing issues. Also we've found that the app works best when using Chrome. Once the app opens you'll need give access to your microphone, and then you can click Record. Once you finish recording all of the words a popup will appear and ask if you'd like to download the data. Simply click OK and the files will be downloaded into the folder from which you are running the flask app (which should be the open-speech-recording folder).

Note: if you want to change the words or counts make sure to kill and re-start the app and open the link in a new incognito window to avoid any cacheing issues at the server or browser level!

You can use the scripts to manipulate the data as follows:

1. You can convert the .ogg files to .wav files using ffmpeg:

sudo apt-get install ffmpeg
mkdir wavs
find *.ogg -print0 | xargs -0 basename -s .ogg | xargs -I {} ffmpeg -i {}.ogg -ar 16000 wavs/{}.wav

2. You can then trim the .wav files with Pete's tool.

mkdir trimmed_wavs
make -C extract_loudest_section/
/tmp/extract_loudest_section/gen/bin/extract_loudest_section 'wavs/*.wav' trimmed_wavs/

3. Finally you can create the directory structure expected by the Tensorflow training script by running another of Pete's scripts adn then compress it into a zip file so it can be easily uploaded to Colab.

python organize_wavs.py
cd output_wavs
zip -r my_dataset.zip *

-Adapted by the Harvard CS249r F2020 team