Naturalistic IN-vehicle Audio Dataset

The NINA dataset is a collection of sounds generated inside and outside (EV sirens) a car cabin. It is intented for research purposes. In this repository, we provide a script to create the dataset.

Sounds are recorded with dashcam or smartphone mic. As recordings are taken in a not controlled environment, we do not have the vehicle speed or the specific recording device model or microphone detail.

Categories:

Class	Clip	Total Duration [sec]
Crash	751	865
Driving	295	1086
Tire skidding	186	208
Horn	261	314
Harsh acceleration	22	63
Talking	265	653
Screaming	157	113
Music	198	821
Pothole	144	138
Meteo (strong rain/hail)	94	3613
Police siren	39	288
Ambulance siren	159	1253
Firetruck siren	76	822

Requirements

In order to run this script, you should have already installed:

• youtube-dl
• sox
• gsed (macOS only)

Files

• dasetCreation.sh: the main script
• youtube_IDs.csv: list of youtube videos.
• labels: folder with txt files, each with the annotation [start time] [end time] [class]

Script running

$ bash datasetCreation.sh ./labels/ ./output

This will create a output folder with a sub-folder per category, including wav files.

Main Youtube channels

• Bad drivers of Italy: https://www.youtube.com/channel/UCqYkaHQFrorRCAj2WgH-G5g/videos
• Car crashes time: https://www.youtube.com/user/CarCrashesTime/videos
• Car crashes time: https://www.youtube.com/channel/UCil5Tyte_KTTrPgt5cC5Q4w/videos

Contribution/Extension

1. Add the new {video_youtube_id} and relative title to the youtube_IDs.csv file.
2. Use Audacity to annotate the file.
   • Open the file
   • Right click on the track -> split stereo to mono and delete one of the two tracks
   • Track menu -> add new track- > label track
   • Select part and them command+b (or ctrl+b) to add the label
   • Edit menu -> Labels -> edit labels -> check and export into a file named {video_youtube_id}.txt
   • Finally to save clips: File menu -> Export -> Export Multiple. Option: split files based on Labels and Name files numbering before Label/track name. At this point file name has this convention {video_youtube_id}_{2_digits_counter}-category (e.g., HRamesEI1Iw_39-ambulance.wav)
3. Save the annotation in a video_youtube_id.txt file in the labels folder.

If you prefere a different tool for annotation (e.g. Elan https://archive.mpi.nl/tla/elan), be sure that the video_youtube_id.txt file is in the format:

starting_time ending_time label_1
starting_time ending_time label_2
...

Wavenet audio generation

nsynth_generate --checkpoint_path=matteo/wavenet-ckpt/model.ckpt-200000 -source_path=matteo/Dataset/AudioFiles/crash/ --save_path=matteo/wavenet_generated/ --batch_size=32 --gpu_number=4

Trimming starting and ending silence from wavenet generated clips:

sox input.wav output.wav silence 1 0.05 1% reverse silence 1 0.05 1% reverse;

Convert Keras model to Tensorflow-lite

tflite_convert --output_file=KfoldNormCNN_3.tflite --keras_model_file=KfoldNormCNN_3.h5

Tensorflow-lite inference on Android

https://thinkmobile.dev/automate-testing-of-tensorflow-lite-model-implementation/