-
Notifications
You must be signed in to change notification settings - Fork 48
Lab 2. Making the Home Sounds Dataset
The data that goes into the neural network is instrumental to what you can get out of it. This lab is focused on helping us to understand the curation of audio data.
As a group, you will all canvas your home environments to capture interesting sounds. This endeavor is the beginning of what we hope will be a larger effort to make a working-from-home equivalent to the Urban Sounds Dataset.
What might be interesting things we'd want to distinguish within your home or apartment? Maybe the sound of different appliances? Maybe the sounds of different cars passing by?
Our goal will to be to collect ~1000 field recordings in ~10 categories, based around a taxonomy we come up with together.
One thing that we will experiment with is using different audio collection tools. Ideally, within each team there will be some people collecting audio with their phones and a few people might have H2N recorders or personal audio recorders. This variety of recording methods will help to introduce some of the artifacts/noise that will help to make subsequent algorithms built on our data robust to variations in recording method.
One thing that will be important will be to record how the data is collected, and what the data is of, so take notes before or after so we can describe the process clearly.
We have set up a Google Drive to upload the data to, but we need to figure out the classification structure in advance rather than just upload a bunch of unlabelled files.
It is always possible to degrade a sound file after it is captured, so we will attempt high quality field recordings: 96kHz/16bit sampling rate, no compression, ideally 10 seconds in length, mono or stereo. (On the iPhone, Settings->search Voice Memos->Audio Quality->Lossless
- Synchronize the time and date on the recorder so that it matches the time and date of your phone.
- We recommend taking a quick reminder photo at each site that audio is recorded. This will help you to remember what the audio clips are of, and will provide a geotag.
- Make sure geotagging is turned on!
Sounds uploaded to the freesound site have the following labelling options, which we feel is worth following as a guide:
- a name
- tags, at least 3 tags
- description
- geotag
An important thing to do before setting out is to record some test clips around your computer, upload the data, and check the sound quality. This helps to make sure that you do not come back from your treks empty-handed, and gives everyone some practice with the overall data capture workflow.
Download the snapshots from the phone and the audio files. The synchronized timestamps for "date created" should help you match the image to the sound file.
Go collect data!
-->NEXT