Voxaboxen

Voxaboxen is a deep learning framework designed to find the start and stop times of (possibly overlapping) sound events in a recording. We designed it with bioacoustics applications in mind, so it accepts annotations in the form of Raven selection tables.

If you use this software in your research, please cite it.

Installation

Requires torch > 2.0 and the corresponding version of torchaudio. Other requirements can be installed with pip install -r requirements.txt.

Once cloned, Voxaboxen can be installed with pip install -e ..

Quick start

Create a train_info.csv file with three columns:

• fn: Unique filename associated with each audio file
• audio_fp: Filepaths to audio files in train set
• selection_table_fp: Filepath to Raven selection tables

Repeat this for the other folds of your dataset, creating val_info.csv and test_info.csv. Run project setup and model training following the template in the Example Usage below.

Notes:

• Audio will be automatically resampled to 16000 Hz mono, no resampling is necessary prior to training.
• Selection tables are .tsv files. We only require the following columns: Begin Time (s), End Time (s), Annotation.

Example usage:

Get the preprocessed Meerkat (MT) dataset:

cd datasets/MT; wget https://storage.googleapis.com/esp-public-files/voxaboxen-demo/formatted.zip; unzip formatted.zip; wget https://storage.googleapis.com/esp-public-files/voxaboxen-demo/original_readme_and_license.md

Project setup:

python main.py project-setup --train-info-fp=datasets/MT/formatted/train_info.csv --val-info-fp=datasets/MT/formatted/val_info.csv --test-info-fp=datasets/MT/formatted/test_info.csv --project-dir=projects/MT_experiment

Train model:

python main.py train-model --project-config-fp=projects/MT_experiment/project_config.yaml --name=demo --lr=.00005 --batch-size=4

Use trained model to infer annotations:

python main.py inference --model-args-fp=projects/MT_experiment/demo/params.yaml --file-info-for-inference=datasets/MT/formatted/test_info.csv

We provide a Colab Notebook with more details about this process.

Evaluation

We trained Voxaboxen on four bioacoustics datasets which represent a variety of animal species and recording conditions. These datasets were chosen because they contain expert annotations of bounding boxes that are precisely aligned with the onsets and offsets of vocalizations. In Table 1 we report the performance of Voxaboxen on a held-out test set from each of these datasets.

As an informal baseline, we fine tuned an image-based Faster-RCNN object detection model on each dataset. Adapted from the Detectron2 code base, these networks were pre-trained with images in the COCO detection task and were fine tuned to detect vocalizations from spectrograms.

For each of these experiments, we performed a small grid search to choose initial learning rate and batch size. For all four datasets, we found that Voxaboxen outperformed Faster-RCNN.

Table 1: Macro-averaged F1 score for each model, dataset pair. To compute these scores, we matched each predicted bounding box with at most one human-annotated bounding box, subject to the condition that the intersection over union (IoU) score of the proposed match was at least 0.5. Two of these datasets (BirdVox 10h and Meerkat) were previously used in the DCASE few-shot detection task. The code for dataset formatting can be found in datasets and the code for replicating these experiments can be found in scripts.

Editing Project Config

After running python main.py project-setup, a project_config.yaml file will be created in the project directory you specified. This config file codifies how different labels will be handled by any model within this project. This config file is automatically generated by the project setup script, but you can edit this file to revise how these labels are handled. There are a few things you can edit:

1. label_set: This is a list of all the label types that a model will be able to output. It is automatically populated with all the label types that appear in the Annotation column of the selection table. If you want your model to ignore a particular label type, perhaps because there are few events with that label type, you must delete that label type from this list.

2. label_mapping: This is a set of key: value pairs. Often, it is useful to group multiple types of labels into one. For example, maybe in your data there are multiple species from the same family, and you would like the model to treat this entire family with one label type. Upon training, Voxaboxen converts each annotation that appears as a key into the label specified by the corresponding value. When modifying label_mapping, you should ensure that each value that appears in label_mapping either also appears in label_set, or is the unknown_label.

3. unknown_label: This is set to Unknown by default. Any sound event labeled with the unknown_label will be treated as an event of interest, but the label type of the event will be treated as unknown. This may be desireable when there are vocalizations that are clearly audible, but are difficult for an annotator to identify to species. When the model is trained, it learns to predict a uniform distribution across possible label types whenever it encounters an event with the unknown_label. When the model is evaluated, it is not penalized for predicting the label of events which are annotated with the unknown_label. The unknown_label should not appear in the label_set.

For example, say you annotate your audio with the labels Red-eyed Vireo REVI, Philidelphia VireoPHVI, and Unsure REVI/PHVI. To reflect your uncertainty about REVI/PHVI, your label_set would include REVI and PHVI, and your label_mapping would include the pairs REVI: REVI, PHVI: PHVI, and REVI/PHVI: Unknown. Alternatively, you could group both types of Vireo together by making your label_set only include Vireo, and your label_mapping include REVI: Vireo, PHVI: Vireo, REVI/PHVI: Vireo.

Other features

Here are some additional options that can be applied during training:

• Flag --stereo accepts stereo audio. Order of channels matters; used for e.g. speaker diarization.
• Flag --bidirectional predicts the ends of events in addition to the beginning, matches starts and ends based on IoU. May improve box regression.
• Flag --segmentation-based switches to a frame-based approach. If used, we recommend putting --rho=1.
• Flag --mixup applies mixup augmentation.

The name

Voxaboxen is designed to put a box around each vocalization (vox). It also rhymes with Roxaboxen.