Hello WildLabs community,

Here's an update on our progress with the AI-based marine mammal sound detection platform.

What We've Accomplished

We've completed our first milestone: a Windows executable for our Burrunan dolphin detection model. This executable allows our AI code to run independently without requiring Python or other dependencies to be installed on the user's machine.

This executable serves as the foundation for our web service - the same packaged code that runs locally will power the online platform, ensuring consistency between our local and web-based offerings.

Current Work

We're currently focused on two main areas:

Web Service Development

We're developing an online platform to make our models accessible via web browsers. The main challenge is finding the right hosting solution that balances cost-effectiveness with handling variable workloads.

Service Optimization

We're developing a more efficient version of our detection algorithms to reduce computational requirements and processing time. This optimisation benefits both the local executables and the planned web service.

Next Steps

Platform Expansion

• Moving the Windows executable from MVP to a production-ready version
• Adding macOS support

Model Integration Approaches

We're evaluating two paths for handling multiple species models:

• Option 1: Create individual executables for each new model we develop. This means every time we complete a new species model, we'll package it as a separate executable.
• Option 2: Build a single executable that can dynamically load different models. This approach is more complex to develop but offers better scalability and could facilitate integration with existing tools like Raven.

Web Service MVP

• Developing a basic browser-based interface for model selection and file upload

Model Library

A system that allows users to choose the species detection model for their acoustic data.

Here are some Burrunan dolphins!

Project Update — Sensors, Sounds, and DIY Solutions (sensores, sonidos y soluciones caseras)

We continue making progress on our bioacoustics project focused on the conservation of Patagonian amphibians, thanks to the support of WILDLABS. Here are some of the areas we’ve been working on in recent months:
(Seguimos avanzando en nuestro proyecto de bioacústica aplicada a la conservación de anfibios patagónicos, gracias al apoyo de WildLabs.Queremos compartir algunos de los frentes en los que estuvimos trabajando estos meses)

1. Hardware

One of our main goals was to explore how to improve AudioMoth recorders to capture not only sound but also key environmental variables for amphibian monitoring. We tested an implementation of the I2C protocol using bit banging via the GPIO pins, allowing us to connect external sensors. The modified firmware is already available in our repository:

👉 https://gitlab.com/emiliobascary/audiomoth

We are still working on managing power consumption and integrating specific sensors, but the initial tests have been promising.

(Uno de nuestros principales objetivos fue explorar cómo mejorar las grabadoras AudioMoth para que registren no sólo sonido, sino también variables ambientales clave para el monitoreo de anfibios. Probamos una implementación del protocolo I2C mediante bit banging en los pines GPIO, lo que permite conectar sensores externos. La modificación del firmware ya está disponible en nuestro repositorio:
https://gitlab.com/emiliobascary/audiomoth
Aún estamos trabajando en la gestión del consumo energético y en incorporar sensores específicos, pero los primeros ensayos son alentadores.)



2. Software (AI)

We explored different strategies for automatically detecting vocalizations in complex acoustic landscapes.

BirdNET is by far the most widely used, but we noted that it’s implemented in TensorFlow — a library that is becoming somewhat outdated.

This gave us the opportunity to reimplement it in PyTorch (currently the most widely used and actively maintained deep learning library) and begin pretraining a new model using AnuraSet and our own data. Given the rapid evolution of neural network architectures, we also took the chance to experiment with Transformers — specifically, Whisper and DeltaNet.

Our code and progress will be shared soon on GitHub.

(Exploramos diferentes estrategias para la detección automática de vocalizaciones en paisajes acústicos complejos. La más utilizada por lejos es BirdNet, aunque notamos que está implementado en TensorFlow, una libreria de que está quedando al margen. Aprovechamos la oportunidad para reimplementarla en PyTorch (la librería de deep learning con mayor mantenimiento y más popular hoy en día) y realizar un nuevo pre-entrenamiento basado en AnuraSet y nuestros propios datos. Dado la rápida evolución de las arquitecturas de redes neuronales disponibles, tomamos la oportunidad para implementar y experimentar con Transformers. Más específicamente Whisper y DeltaNet. Nuestro código y avances irán siendo compartidos en GitHub.)

3. Miscellaneous

Alongside hardware and software, we’ve been refining our workflow.

We found interesting points of alignment with the “Safe and Sound: a standard for bioacoustic data” initiative (still in progress), which offers clear guidelines for folder structure and data handling in bioacoustics. This is helping us design protocols that ensure organization, traceability, and future reuse of our data.

We also discussed annotation criteria with @jsulloa to ensure consistent and replicable labeling that supports the training of automatic models.

We're excited to continue sharing experiences with the Latin America Community— we know we share many of the same challenges, but also great potential to apply these technologies to conservation in our region.

(Además del trabajo en hardware y software, estamos afinando nuestro flujo de trabajo. Encontramos puntos de articulación muy interesantes con la iniciativa “Safe and Sound: a standard for bioacoustic data” (todavía en progreso), que ofrece lineamientos claros sobre la estructura de carpetas y el manejo de datos bioacústicos. Esto nos está ayudando a diseñar protocolos que garanticen orden, trazabilidad y reutilización futura de la información. También discutimos criterios de etiquetado con @jsulloa, para lograr anotaciones consistentes y replicables que faciliten el entrenamiento de modelos automáticos. Estamos entusiasmados por seguir compartiendo experiencias con Latin America Community , con quienes sabemos que compartimos muchos desafíos, pero también un enorme potencial para aplicar estas tecnologías a la conservación en nuestra región.)