The technology

We find marine species vocalizations using state-of-the-art acoustic AI technology.
Searching for vocalisations in a sea of background noise is tedious and time-consuming. However, our models could save you time and effort.

How do we do it?

We have a framework for training deep-learning models suitable for acoustic detection and classification. Our experienced research and development team keeps improving our capabilities to ensure we’re always aligned with the state-of-the-art and best practices of acoustic classification.

We have models trained for various marine species:

Humpback whales

Killer Whale

Codfish

Harbour porpoise

Common dolphin

Southern resident killer whales

False killer whales

Antillean Manatees

North Atlantic right whale

What makes us unique?

Our people are important to us much as the whales!

Robustness Through Noise Suppression

Our models incorporate sophisticated signal processing to isolate target vocalisations even within complex, noisy soundscapes. This ensures data integrity regardless of environmental conditions or equipment limitations.

Adaptive Learning for Precision

Transfer learning and fine-tuning methodologies allow models to be rapidly adapted to new environments or species. To detect new species, we only need a few hundred annotations of existing sounds.

Speed

Our models are really quick, allowing us to review acoustic data 400 times as fast as a human. This also means that our models can run in real time.

Interdisciplinary Expertise

Our team combines a deep understanding of acoustic principles, real-world signal acquisition experience, and cutting-edge machine learning techniques. This ensures solutions that effectively address the challenges of diverse acoustic environments.

Modular Design for Customization

Our framework utilises interchangeable audio representations, processing algorithms, and model types. This enables us to tailor pipelines precisely to your specific dataset and analysis requirements.

Scalability for Big Data

Our solutions are architected for distributed processing, gracefully handling large-scale acoustic datasets. Analys vast amounts of data without compromising efficiency or performance.

human in the loop - for better performance

In Deep Voice, we value human knowledge and capabilities and wish to enjoy human-machine interaction to the fullest. We provide active learning mechanisms for more efficient human annotation of training data to reduce the annotations needed and ensure they provide the best possible result. Also, it’s possible to collaborate with our annotation team to aid in the annotation and data preparation process for the models.

How is a model trained?

We use a human-labeled dataset to train our model for animal classification. Using those annotations, the models are trained to extract the patterns and information hidden in the data and learn to build a strong understanding of the classification domain and signal uniqueness. The models also learn to generalise and find calls that the human eye and hearing might miss.  

For more technical information, please refer to our paper describing the training and usage of our models

OUR SERVICES

Automatic detection

Automatically detect animal sounds x400 faster than a human.

Automatic classification

Distinguish between different sound types made in a single recording

Real-time annotation

Indicating relevant noises in real time. This feature is still in work.

Tagging services

Our expert team can help you manually annotate data, regardless of our AI services.

The cost

Being staffed by volunteers and with the goal of helping organizations and lawmakers make data-driven decisions to protect animals and their surroundings, we offer our services at negligible cost (a few hundred dollars) Since we genuinely believe in our goal, we never let monetary issues prevent us from helping to better the world. Note our tagging services are paid separately with the money paid for our experts’ time.