DeepForestSound: a multi-species automatic detector for passive acoustic monitoring in African tropical forests, a case study in Kibale National Park

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "DeepForestSound: a multi-species automatic detector for passive acoustic monitoring in African tropical forests, a case study in Kibale National Park".

Jane: Here's the gist of the paper. DeepForestSound (DFS) is a multi-species automatic detection model designed for passive acoustic monitoring in African tropical forests, which outperforms existing tools for non-avian taxa by leveraging task-oriented training data and parameter-efficient fine-tuning techniques.

Tom: Passive Acoustic Monitoring (PAM) is a powerful alternative for large-scale and long-term biodiversity assessment in tropical rainforests By recording environmental sounds, PAM enables the detection and monitoring of vocalizing species such as birds, primates, and elephants with minimal disturbance Traditional field-based survey methods are often costly, time-consuming, invasive, and limited in spatial and temporal coverage or can be dangerous in dense African rainforests Recent advances in deep learning have significantly improved ecoacoustic analysis, outperforming traditional signalprocessing methods in many tasks However, general-purpose models often struggle when biodiversity datasets lack sufficient coverage in specific geographic regions or for rare species Transfer learning using pretrained audio models has emerged as a promising strategy to address the scarcity of labeled data in contexts like African rainforests Data and Training Pipeline The study utilized a semi-supervised pipeline to generate labeled examples from unannotated recordings The pipeline consists of four main steps: (i) detection of candidate vocalizations based on energy criteria, (ii) extraction of embeddings using pretrained models, (iii) clustering in a reduced feature space, and (iv) manual validation of clusters to produce labeled training data Candidate segments were detected based on dB RMS energy within literature-derived frequency bands and expected duration ranges for each species Embeddings were obtained using embeddings from the last layer of three pretrained models used separately to capture complementary acoustic representations, namely AST, Perch v2, and BirdNET Dimensionality reduction and clustering were achieved with UMAP and HDBSCAN respectively to extract species-specific vocalizations Model Architecture and Fine-Tuning The core model, DFS, relies on parameter-efficient fine-tuning of the Audio Spectrogram Transformer (AST) backbone using Low-Rank Adaptation (LoRA) This approach injects trainable low-rank matrices into the attention layers of the pretrained model while keeping the original parameters frozen, drastically reducing trainable parameters from 88.30M to 1.04M A baseline variant, DFS-Linear, uses the same pretrained AST encoder but keeps all backbone parameters frozen and trains only a single linear classification layer on top of the extracted embeddings The model was trained for up to eight epochs on a laptop equipped with an NVIDIA RTX 3080 GPU Data augmentation strategies applied included time masking and frequency masking following the SpecAugment paradigm Additionally, a mixup-based augmentation strategy was employed to mitigate class imbalance by uniformly sampling the energy ratio between positive and background signals between 3 and 15 dB Performance Evaluation Model performances were assessed using Average Precision (AP)

35: and the best F1 score, defined as the harmonic mean of precision and recall The models were evaluated on the Sebitoli 2025 dataset, consisting of 1,200 manually annotated one-minute recordings collected at locations different from the training data DFS achieves the highest best F1-scores and AP values for the majority of evaluated species (8 out of 12), particularly for non-avian taxa Across all primate species, DFS reaches a mean best F1 of 0.932 and a mean AP of 0.964 For birds, DFS achieves an average best F1 of 0.856 and an AP of 0.896 Perch v2 demonstrates strong performance on several bird species, achieving the highest per-species scores among baseline methods for avian taxa However, its performance on non-avian species remains substantially lower DFS is designed to complement large-scale generalist ecoacoustic models rather than substitute them for broad biodiversity assessments The results demonstrate that task-oriented training data and domain-specific segmentation substantially improve detection performance in acoustically complex tropical environments The ability of DFS to accurately detect threatened species is relevant for conservation-oriented monitoring programs, as it achieves the highest detection performance for four of the six listed threatened taxons This work highlights that in data-scarce ecological contexts, task-specific models trained on curated local datasets can outperform large-scale generalist models Future work will focus on extending the framework to additional species and forests The code and pretrained models for DFS are publicly available at https://github.com/MNHN-OFVI/DFS--multi-species-automatic-detector-for-pam-in-african--tropical--forests Page 1 --- DeepForestSound: a multi-species automatic detector for passive acoustic monitoring in African tropical forests, a case study in Kibale National Park Gabriel Dubus∗†, Theau d’Audiffret ´ Claire Auger∗†, Raphael Cornette ¨ Sylvain Haupert¶, Innocent Kasekendi† Raymond Katumba† Hugo Magaldi∗† Lise Pernel∗∥ Harold Rugonge† Jer´ ome Sueur ˆ ¶ John Justice Tibesigwa‡ Sabrina Krief∗† Eco-Anthropologie, Museum National d’Histoire Naturelle, UMR7206, Paris, France ´ Sebitoli Chimpanzee Project, Great Ape Conservation Project, Sebitoli, Kibale National Park Fort Portal Uganda † Uganda Wildlife Authority Kampala Uganda § N’lab Nitidae Association Montpellier France ¶ Centre d’Ecologie et des Sciences de la Conservation (CESCO) UMR CNRS Sorbonne Universite 7204 Museum national d’Histoire Naturelle Paris France ∥ Institut de Systematique, Evolution, Biodiversité (ISYEB) UMR7205 Centre National de la Recherche Scientifique MNHN SU EPHE-PSL UA Paris France Page 2 --- Tropical rainforests host a large proportion of terrestrial biodiversity, yet they are among the most threatened ecosystems due to habitat loss, climate change, and increasing human pressure Effective biodiversity monitoring in these environments is therefore a major conservation priority Passive Acoustic Monitoring (PAM) has emerged as a powerful alternative for large-scale and long-term biodiversity assessment By recording environmental sounds, PAM enables the detection and monitoring of vocalizing species such as birds

4: , primates

5: , and elephants

6: with minimal disturbance Despite these advantages, the large volume of recordings and the high acoustic complexity of soundscapes pose significant challenges for data analysis Recent advances in deep learning have significantly improved ecoacoustic analysis, outperforming traditional signalprocessing methods in many tasks

8: ,

9: Several large-scale models have been developed for general-purpose audio classification as well as for biodiversity applications, including the AST

10: , BirdNET

11: , and Perch v2

12: However, their effectiveness can be limited in certain contexts In particular, biodiversity datasets often lack sufficient coverage in specific geographic regions or for rare species, leading to limited applicability in biodiversity monitoring For example, iNaturalist contains far fewer recordings from African regions compared to Europe

13: , and BirdNET has been shown to perform less accurately in Africa or Oceania compared to Europe or North America

16: To address the scarcity of labeled data in such contexts, transfer learning using pretrained audio models has emerged as a promising strategy This approach is particularly relevant in African rainforests, where high species diversity contrasts sharply with the limited availability of labeled acoustic data <ref:2604.

Title and authors: Tom: So we’re looking at the paper DeepForestSound: a multi-species automatic detector for passive acoustic monitoring in African tropical forests, a case study in Kibale National Park. What’s the main takeaway about what they built and why it matters?

Jane: Basically, this paper is about an automatic detection system for passive acoustic monitoring in those tough African tropical forests where getting labeled data is really hard. They introduce DeepForestSound, or DFS, which uses a semi-supervised workflow to generate its own training data so it can learn to find different species sounds.

Lu: The core idea there is using transfer learning with pre-trained audio models to adapt them for these specific tasks without needing massive amounts of labeled acoustic data. They specifically focused on the scarcity of labeled data in these regions, which is a major problem for general models.

Meng: So, the system isn't just one big model trying to do everything; it’s designed to be task-oriented and leverage existing knowledge from other audio models. What kind of specific improvements did they make to their approach?

Tom: They focused on parameter-efficient fine-tuning using Low-Rank Adaptation, or LoRA, which lets them adapt the Audio Spectrogram Transformer backbone while keeping most of the original parameters frozen. This drastically cuts down the trainable parameters from eighty-eight point three million down to one point zero four million <ref:2604.08087#pg1>.

Jane: That reduction in size is important because it makes fine-tuning much more manageable when you are working with limited labeled data, which is exactly the scenario they are tackling. They also used a dual pre-processing configuration to handle different acoustic characteristics, using low-frequency settings for things like elephant rumbles and mid-frequency settings for bird sounds <ref:2604.08087#pg3>.

Lu: That frequency-dependent attenuation they applied, which simulates vegetation scattering and absorption in decibels as a function of frequency, moves the method beyond just standard data augmentation to incorporate environmental realism <ref:2604.08087#pg2>.

Meng: I’m curious about the deployment side—if we take that small parameter count, does it still run efficiently when we try to deploy this at scale across a massive monitoring network? What breaks first in that setup?

Tom: They also have a baseline variant called DFS-Linear, which keeps the entire backbone frozen and just trains a single linear classification layer on top of the extracted embeddings. It’s a direct comparison point to see if the complexity of LoRA is actually worth it for this specific task <ref:2604.08087#pg1>.

Jane: And when they tested it, they looked at the Sebitoli two thousand twenty-five dataset, which consisted of one thousand two hundred manually annotated recordings collected at locations different from where they trained their model <ref:2604.08087#pg3>. DFS achieved the highest best F1-scores and AP values for most of the species they tested.

Lu: Specifically, across all primate species, DFS reaches a mean best F1 of zero point nine three two and a mean AP of zero point nine six four <ref:2604.08087#pg3>. For birds, the average best F1 is zero point eight five six and an AP of zero point eight nine six <ref:2604.08087#pg3>.

Meng: So it performs very well on primates, which makes sense given the data they used for that part of the training. But what about the non-avian taxa? Where does it fall off compared to those established models?

Tom: It still shows strong performance for eight out of twelve species evaluated across all primate groups. They note that Perch v2 is very strong on some bird species, achieving the highest per-species scores among baseline methods for avian taxa, but its performance on non-avian species remains substantially lower <ref:2604.08087#pg3>.

Jane: The implication here is that task-oriented training data and domain-specific segmentation really do improve detection performance in acoustically complex tropical environments, which is what the authors are stressing with their work on DeepForestSound.

Lalam: I’m confident in this finding because the structure of their semi-supervised pipeline directly addresses the data scarcity issue they identified at the start <ref:2604.08087#pg2>.

Tom: So, to wrap up, this DeepForestSound model is designed to complement large-scale generalist ecoacoustic models rather than replace them entirely for broad biodiversity assessments.

Jane: Exactly. It shows that when you have limited labeled data in a place like an African tropical forest, task-specific models trained on curated local datasets can actually perform better than the big generalist models.

Lu: For future work, they plan to extend this framework to additional species and forests, which suggests the core representation learning idea has real potential for broader ecological application <ref:2604.08087#pg1>.

Meng: From a deployment standpoint, it’s a solid tool for conservation monitoring because it targets threatened taxons, achieving the highest detection performance for four of the six listed threatened taxa <ref:2604.08087#pg3>.

Tom: So we've looked at how they used LoRA to adapt an AST backbone, how they built a semi-supervised pipeline from energy detection to manual validation, and what the measured F1 scores look like on their Sebitoli two thousand twenty-five dataset <ref:2604.08087#pg3>.

Jane: It’s a case study in moving toward more specialized AI solutions for biodiversity monitoring in data-scarce regions. That’s what DeepForestSound is all about.

The paper's summary: Tom: So, to recap, this paper is about DeepForestSound using LoRA fine-tuning and a multi-model embedding clustering pipeline to detect multiple species in tropical forests with good results for primates and threatened taxa.

Jane: Right, so basically they took existing audio models and made them work better for these specific African forest sounds by training them on data they helped create themselves.

Lu: The core idea is using that pre-trained model but making it task-specific without training all eight hundred million parameters from scratch. It’s about finding the right representation for those complex forest sounds.

Meng: That reduction in trainable parameters from eighty-eight point three million down to one point zero four million, which makes the fine-tuning process much more manageable when you’re dealing with limited labeled data in a specific geographic area.

Tom: I see that parameter count as good for lower computational overhead for running this model out in the field compared to massive generalist models. But I wonder if that low parameter count sacrifices some of the accuracy we might need for rare species detection.

Jane: They also used embeddings from three different pretrained models—AST, Perch v2, and BirdNET—to capture those complementary acoustic representations before clustering them with UMAP and HDBSCAN to find those species-specific vocalizations.

Lu: That multi-model embedding extraction is smart because it gives the final classification layer a richer set of acoustic features to work with than just one model could provide on its own.

Meng: I’m still thinking about the augmentation strategy; they used time masking and frequency masking, plus they simulated forest effects using frequency-dependent attenuation in dB as a function of frequency. That environmental realism is something I'd want to see tested on real traffic data at scale.

Tom: The results showed DFS achieving the highest best F1-scores and AP values for most of the evaluated species, particularly for non-avian taxa, reaching a mean best F1 of zero point nine three two for primates.

Jane: That means it’s doing very well with those primates, which is important because those are often easier to label initially. But they also showed that DFS can detect threatened taxons better than other methods.

Lu: The main implication here is that task-oriented training data and domain-specific segmentation substantially improve detection performance in acoustically complex tropical environments when you have limited labeled acoustic data.

Meng: So for a conservation program, this means they could use this to monitor specific threatened species, even if the overall model isn't perfect for everything else.

Tom: It’s a case study in how to build specialized tools that outperform generalist models when the data context is very specific.

Jane: This DeepForestSound paper shows us how we can leverage transfer learning with parameter-efficient fine-tuning to get high performance where labeled data is scarce, especially in places like African rainforests.

Lu: It suggests that focusing on curated local datasets for task-specific models can outperform large scale generalist models in these kinds of ecological contexts.

Meng: If you’re deploying something like this, the key will be figuring out the cost and complexity of running that LoRA adaptation efficiently on edge hardware.

Lalam: I’m fairly sure this work is very sure about its findings because they focused on creating a robust pipeline from unannotated audio to a trained detector for specific species.

Tom: So, to recap, DeepForestSound uses LoRA fine-tuning and a multi-model embedding clustering pipeline to detect multiple species in tropical forests with good results for primates and many threatened taxa.

Jane: It’s really about showing how targeted AI can solve problems where the data just isn't there otherwise.

Lu: The next step for this kind of research would be extending the framework to more species and different forest types.

The paper's improvements: Tom: So we’ve looked at DeepForestSound, this multi-species automatic detector for passive acoustic monitoring in African tropical forests, and what I see is that they used a parameter-efficient fine-tuning approach with LoRA to adapt an Audio Spectrogram Transformer backbone.

Jane: Right, so the core idea is using that pre-trained model but making it task-specific without training all eight hundred million parameters from scratch. It’s about finding the right representation for those complex forest sounds.

Lu: Exactly, they reduced the trainable parameters from eighty-eight point three million down to one point zero four million, which makes the fine-tuning process much more manageable when you’re dealing with limited labeled data in a specific geographic area.

Meng: From a deployment standpoint, that small parameter count is good; it means lower computational overhead for running this model out in the field compared to massive generalist models. But I wonder if that low parameter count sacrifices some of the accuracy we might need for rare species detection.

Lalam: I’m fairly certain this approach is solid because they used a semi-supervised pipeline where they detected candidate vocalizations based on energy criteria and then manually validated those clusters to create the training data.

Tom: That part—the detection of candidates based on dB RMS energy within literature-derived frequency bands—is how they turned raw audio into usable labels for their model. It’s a very structured way to approach unlabeled data.

Jane: And they used embeddings from three different pretrained models—AST, Perch v2, and BirdNET—to capture those complementary acoustic representations before clustering them with UMAP and HDBSCAN to find those species-specific vocalizations.

Lu: That multi-model embedding extraction is smart because it gives the final classification layer a richer set of acoustic features to work with than just one model could provide on its own.

Meng: I’m still thinking about the augmentation strategy; they used time masking and frequency masking, plus they simulated forest effects using frequency-dependent attenuation in dB as a function of frequency. That environmental realism is something I'd want to see tested on real traffic data at scale.

Tom: The results showed DFS achieving the highest best F1-scores and AP values for most of the evaluated species, particularly for non-avian taxa, reaching a mean best F1 of zero point nine three two for primates.

Jane: That means it’s doing very well with those primates, which is important because those are often easier to label initially. But they also showed that DFS can detect threatened taxons better than other methods.

Lu: The main implication here is that task-oriented training data and domain-specific segmentation substantially improve detection performance in acoustically complex tropical environments when you have limited labeled acoustic data.

Meng: So for a conservation program, this means they could use this to monitor specific threatened species, even if the overall model isn't perfect for everything else.

Tom: It’s a case study in how to build specialized tools that outperform generalist models when the data context is very specific.

Jane: This DeepForestSound paper shows us how we can leverage transfer learning with parameter-efficient fine-tuning to get high performance where labeled data is scarce, especially in places like African rainforests.

Lu: It suggests that focusing on curated local datasets for task-specific models can outperform large scale generalist models in these kinds of ecological contexts.

Meng: If you’re deploying something like this, the key will be figuring out the cost and complexity of running that LoRA adaptation efficiently on edge hardware.

Lalam: I’m fairly sure this work is very sure about its findings because they focused on creating a robust pipeline from unannotated audio to a trained detector for specific species.

Tom: So, to recap, DeepForestSound uses LoRA fine-tuning and a multi-model embedding clustering pipeline to detect multiple species in tropical forests with good results for primates and many threatened taxa.

Jane: It’s really about showing how targeted AI can solve problems where the data just isn't there otherwise.

Lu: The next step for this kind of research would be extending the framework to more species and different forest types.

Conclusion: Tom: So we’ve looked at DeepForestSound, this multi-species automatic detector for passive acoustic monitoring in African tropical forests, and what I see is that they used a parameter-efficient fine-tuning approach with LoRA to adapt an Audio Spectrogram Transformer backbone.

Jane: Right, so the core idea is using that pre-trained model but making it task-specific without training all eight hundred million parameters from scratch. It’s about finding the right representation for those complex forest sounds.

Lu: Exactly, they reduced the trainable parameters from eighty-eight point three million down to one point zero four million, which makes the fine-tuning process much more manageable when you’re dealing with limited labeled data in a specific geographic area.

Meng: From a deployment standpoint, that small parameter count is good; it means lower computational overhead for running this model out in the field compared to massive generalist models. But I wonder if that low parameter count sacrifices some of the accuracy we might need for rare species detection.

Lalam: I’m fairly certain this approach is solid because they used a semi-supervised pipeline where they detected candidate vocalizations based on energy criteria and then manually validated those clusters to create the training data.

Tom: That part—the detection of candidates based on dB RMS energy within literature-derived frequency bands—is how they turned raw audio into usable labels for their model. It’s a very structured way to approach unlabeled data.

Jane: And they used embeddings from three different pretrained models—AST, Perch v2, and BirdNET—to capture those complementary acoustic representations before clustering them with UMAP and HDBSCAN to find those species-specific vocalizations.

Lu: That multi-model embedding extraction is smart because it gives the final classification layer a richer set of acoustic features to work with than just one model could provide on its own.

Meng: I’m still thinking about the augmentation strategy; they used time masking and frequency masking, plus they simulated forest effects using frequency-dependent attenuation in dB as a function of frequency. That environmental realism is something I'd want to see tested on real traffic data at scale.

Tom: The results showed DFS achieving the highest best F1-scores and AP values for most of the evaluated species, particularly for non-avian taxa, reaching a mean best F1 of zero point nine three two for primates.

Jane: That means it’s doing very well with those primates, which is important because those are often easier to label initially. But they also showed that DFS can detect threatened taxons better than other methods.

Lu: The main implication here is that task-oriented training data and domain-specific segmentation substantially improve detection performance in acoustically complex tropical environments when you have limited labeled acoustic data.

Meng: So for a conservation program, this means they could use this to monitor specific threatened species, even if the overall model isn't perfect for everything else.

Tom: It’s a case study in how to build specialized tools that outperform generalist models when the data context is very specific.

Jane: This DeepForestSound paper shows us how we can leverage transfer learning with parameter-efficient fine-tuning to get high performance where labeled data is scarce, especially in places like African rainforests.

Lu: It suggests that focusing on curated local datasets for task-specific models can outperform large scale generalist models in these kinds of ecological contexts.

Meng: If you’re deploying something like this, the key will be figuring out the cost and complexity of running that LoRA adaptation efficiently on edge hardware.

Lalam: I’m fairly sure this work is very sure about its findings because they focused on creating a robust pipeline from unannotated audio to a trained detector for specific species.

Tom: So, to recap, DeepForestSound uses LoRA fine-tuning and a multi-model embedding clustering pipeline to detect multiple species in tropical forests with good results for primates and many threatened taxa.

Jane: It’s really about showing how targeted AI can solve problems where the data just isn't there otherwise.

Lu: The next step for this kind of research would be extending the framework to more species and different forest types.

Meng: And for me, it means if you have a specific conservation goal, you build a focused tool like this instead of trying to build one massive model that tries to do everything poorly.

Lalam: We’ll see what they come up with next when they try to apply this framework to a different type of forest or species.