Voogle

Voogle is an audio search engine that uses vocal imitations of the desired sound as the search query.

Voogle is built in Python 3.6 and Javascript, using Node.js. Voogle runs best in Google Chrome.

Installation

Server

Voogle backend dependencies are installed with pip install -r requirements.txt.

Note: Windows and Linux users must have FFmpeg installed.

Interface

Voogle frontend dependencies are installed with npm install.

Note: You must have Node.js installed before you can run npm install.

Available Datasets

Any collection of audio files can be used as sounds returned by Voogle in response to a vocal query. The Interactive Audio Lab has released 2 datasets specifically for the training of query-by-vocal-imitation models: Vocal Imitation Set and VocalSketch [1, 2]. A small test dataset for demos can be downloaded here.

Audio files should be placed in data/audio/<dataset_name>. The dataset used during execution can be specified in config.yaml.

Available Models

Interactive Audio Lab has released the following models for query-by-vocal-imitation:

• siamese-style: a siamese-style neural network [3]
  • weight file
• VGGish-embedding: cosine similarity of VGGish embeddings [4]
  • weight file
• mcft: multi-resolution common-fate transform [5]

Weight files should be placed in model/weights. The model used during execution can be specified in config.yaml.

Setup

After installing the dependencies, a dataset, and a model, the Voogle app can be deployed.

Deploying Locally

1. Start the server by running npm run production.
2. Navigate to localhost:5000 in your browser.

From there, please follow the directions found under "Show Instructions". Enjoy!

Note: There are currently two frontend interfaces available for Voogle. If you would like to use the alternate interface, use the command npm run old-interface instead during step 1.

Testing

Unit tests can be run with npm run test.

Extending

Voogle can be extended to incorporate additional models and datasets. If you would like to make your model or dataset available to all users of Voogle, contact interactiveaudiolab@gmail.com.

Adding a model

• Define your model as a subclass of QueryByVoiceModel with all abstract methods implemented as described.
• Add the model constructor to factory.py.
• Place your model's weights in model/weights.
• Update the model name and filepath in config.yaml.

An example model can be found here.

Adding a dataset

• Define your dataset as a subclass of QueryByVoiceDataset with all abstract methods implemented as described.
• Add the dataset constructor to factory.py.
• Place the audio files in data/audio/<dataset_name>.
• Update the dataset name in config.yaml.

An example dataset can be found here.

References

• [1] Bongjun Kim, Madhav Ghei, Bryan Pardo, and Zhiyao Duan, "Vocal Imitation Set: a dataset of vocally imitated sound events using the AudioSet ontology," Proceedings of the Detection and Classification of Acoustic Scenes and Events 2018 Workshop (DCASE2018), Surrey, UK, Nov. 2018. [paper link]
• [2] Mark Cartwright and Bryan Pardo, "Vocalsketch: Vocally imitating audio concepts," Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (ACM), 2015. [paper link]
• [3] Yichi Zhang, Bryan Pardo, and Zhiyao Duan, "Siamese Style Convolutional Neural Networks for Sound Search by Vocal Imitation," IEEE/ACM Transactions on Audio Speech and Language Processing. [paper link]
• [4] Bongjun Kim and Bryan Pardo, "Improving Content-based Audio Retrieval by Vocal Imitation Feedback," IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2019.
• [5] Fatemeh Pishdadian and Bryan Pardo. “Multi-resolution Common Fate Transform,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2018. [paper link]