Chimpanzee Vocalization CNN10 (trained with synthetic data)

A 10-layer Convolutional Neural Network for detecting chimpanzee vocalizations in bioacoustic recordings. Part of the animal-sounds pipeline developed at Utrecht University.

Model description

• Architecture: CNN10 — 4 convolutional blocks (64→128→256→512 channels), global average pooling, fully-connected head, 2-class output (chimpanzee vocalization vs. background)
• Input: 3-channel mel spectrogram representation (mel spectrogram → PCEN normalization → per-file z-normalization → delta features)
• Preprocessing: recordings resampled to 48,000 Hz, filtered with a Butterworth bandpass filter (100–2,000 Hz)
• Frame length: 0.5-second audio frames

This checkpoint was trained with synthetic data: real chimpanzee vocalizations and background recordings from Cameroon, augmented with synthetic examples created by mixing Cameroonian chimpanzee calls with Zambian background noise, to improve robustness in unseen acoustic environments.

A version trained without synthetic data is also available: utrechtuniversity/chimp-vocalization-cnn10-sanctuary.

Training data

• Sanctuary recordings: Mefou Primate Sanctuary, Cameroon (December 2019 – January 2020), publicly available (Zwerts et al., 2024)
• Synthetic augmentation: Cameroonian chimpanzee calls mixed with background noise from Chimfunshi Wildlife Orphanage Trust (CWOT), Zambia

Evaluation

Evaluated at file level on held-out recordings from Chimfunshi Wildlife Orphanage Trust (CWOT), Zambia — a different acoustic environment from the training data, used to assess cross-environment generalization. Two recorders ("13b" and "14a") were evaluated separately.

Metrics are for the positive (chimpanzee) class unless labeled "-macro" (averaged across both classes).

Usage

This model is not integrated with transformers — it's a plain PyTorch model exported with PyTorchModelHubMixin. Copy the class definition below, then load the trained weights directly from the Hub.

import torch.nn as nn
from huggingface_hub import PyTorchModelHubMixin


def _cnn_block(in_ch, out_ch):
    return nn.Sequential(
        nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
        nn.BatchNorm2d(out_ch),
        nn.ReLU(inplace=True),
    )


class CNN10Hub(nn.Module, PyTorchModelHubMixin):
    def __init__(self, num_channels=3, num_labels=2, dropout_rate=0.5, **kwargs):
        super().__init__()
        layers = [
            _cnn_block(num_channels, 64), _cnn_block(64, 64), nn.AvgPool2d(2), nn.Dropout(dropout_rate),
            _cnn_block(64, 128), _cnn_block(128, 128), nn.AvgPool2d(2), nn.Dropout(dropout_rate),
            _cnn_block(128, 256), _cnn_block(256, 256), nn.AvgPool2d(2), nn.Dropout(dropout_rate),
            _cnn_block(256, 512), _cnn_block(512, 512), nn.AvgPool2d(2), nn.Dropout(dropout_rate),
            nn.AdaptiveAvgPool2d(1), nn.Dropout(dropout_rate), nn.Flatten(),
            nn.Linear(512, 512), nn.ReLU(), nn.Dropout(dropout_rate),
            nn.Linear(512, num_labels),
        ]
        self.acoustic_model = nn.Sequential(*layers)

    def forward(self, x):
        return self.acoustic_model(x)


model = CNN10Hub.from_pretrained("utrechtuniversity/chimp-vocalization-cnn10-synthetic")
model.eval()

Important: the model expects input tensors matching the exact preprocessing pipeline described above (3-channel mel + PCEN + delta features, 48kHz source audio, 0.5s frames). Raw audio or standard mel spectrograms without this preprocessing will produce incorrect predictions.

A standalone preprocess.py (dependent only on librosa, numpy, scipy — no need to install the full source package) is included in this repo, reproducing the exact training-time feature extraction:

from huggingface_hub import hf_hub_download
import importlib.util
import torch

path = hf_hub_download("utrechtuniversity/chimp-vocalization-cnn10-synthetic", "preprocess.py")
spec = importlib.util.spec_from_file_location("preprocess", path)
preprocess = importlib.util.module_from_spec(spec)
spec.loader.exec_module(preprocess)

features = preprocess.extract_features("my_recording.wav")   # (3, n_mel, n_frames)
chunks = preprocess.chunk_features(features)                 # (n_chunks, 3, n_mel, 64)

with torch.no_grad():
    probs = torch.softmax(model(torch.from_numpy(chunks).float()), dim=1)

See the feature extraction source in the GitHub repository for further background on the pipeline.

Intended use and limitations

• Trained specifically for chimpanzee vocalization detection at the sanctuary/wildlife-orphanage recording conditions described above. Performance on substantially different recording setups, microphone types, or species is not guaranteed.
• Although developed for chimpanzee vocalizations, the underlying pipeline in the source repository can be adapted to other species given labeled training data.

License

Code released under Apache 2.0, matching the source GitHub repository.

Citation

The chimpanzee vocalization dataset used to train this model:

Zwerts, J. A., Treep, J., Kaandorp, C. S., Meewis, F., Koot, A. C., & Kaya, H. (2021).
Introducing a central african primate vocalisation dataset for automated species classification.
arXiv preprint. https://arxiv.org/pdf/2101.10390.pdf

Zwerts, J., Treep, J., Zahedi, P., & Kaandorp, C. (2024).
Central African Primate vocalization bioacoustics dataset: Yoda Data publication platform of Utrecht University.

A paper describing the CNN10/CNN12 models and cross-environment evaluation presented here is currently in preparation. This model card will be updated with the correct citation once available — in the meantime, please contact Dr. Joeri A. Zwerts (j.a.zwerts@uu.nl) for citation guidance.

Links

• Source code and full pipeline: github.com/UtrechtUniversity/animal-sounds
• Contact: Dr. Joeri A. Zwerts, Assistant Professor — j.a.zwerts@uu.nl
• Technical/repository support: Research Engineering team, Utrecht University — research.engineering@uu.nl