This model detects the presence of Southern Resident Killer Whale (SRKW) calls in audio recordings from hydrophone networks. It was trained on labeled audio data from the Orcasound hydrophone network in Puget Sound, Washington.

Architecture: ResNet50 with custom classification head Input: Mel spectrogram (1 channel, 256 mel bins, 312 time frames) Output: Binary classification (orca call present / not present) Framework: PyTorch (ported from FastAI)

Model Details

Developed by: Akash Mahajan, Prakruti Gogia, Aayush Agrawal
Model type: Audio classification (binary)
License: OrcaHello RAIL (Responsible AI License) - See LICENSE file
Finetuned from: ResNet50 ImageNet pretrained weights
Date trained: November 2020
Code: Orcasound/aifororcas-livesystem/InferenceSystem

Note: This is a port of the original FastAI model to pure PyTorch for easier usage and deployment.

Usage

Follow setup instructions in the code linked above. Then use module src as follows below:

Quick Start: Detection in audio file

Internally handles audio pre-processing into segments and batched inference.

python

from src.model_v1 import OrcaHelloSRKWDetectorV1

model = OrcaHelloSRKWDetectorV1.from_pretrained("orcasound/orcahello-srkw-detector-v1")
result = model.detect_srkw_from_file("audio.wav")

print(f"Orca detected: {result.global_prediction}")
print(f"Confidence: {result.global_confidence:.1f}%")
print(f"Segment predictions: {result.segment_predictions}")

Advanced: Per-segment processing

For fine-grained control over audio preprocessing and per-segment predictions.

python

from src.model_v1 import OrcaHelloSRKWDetectorV1, DetectorInferenceConfig
from src.model_v1.audio_frontend import AudioPreprocessor

# Load configuration
config = DetectorInferenceConfig.from_yaml("config.yaml")
model = OrcaHelloSRKWDetectorV1.from_pretrained("orcasound/orcahello-srkw-detector-v1")

# Process segments manually
preprocessor = AudioPreprocessor(config)
for mel_spec, start_s, duration_s in preprocessor.process_segments("audio.wav"):
    X = mel_spec.unsqueeze(0).to(model.device)
    confidence = model.predict_call(X)
    print(f"Segment at {start_s:.1f}s: confidence={confidence:.3f}")

Configuration

Configuration can be loaded from YAML to modify inference behavior:

YAML format (config.yaml):

yaml

audio:
  downmix_mono: true
  resample_rate: 20000

spectrogram:
  sample_rate: 16000
  n_fft: 2560
  hop_length: 256
  mel_n_filters: 256
  mel_f_min: 0.0
  mel_f_max: 10000.0

inference:
  window_s: 2.0               # segment length
  window_hop_s: 1.0           # hop between segments
  max_batch_size: 8           # max segments to process at once in detect_srkw_from_file
  strict_segments: true       # if false, allow partial final segment

global_prediction:
  aggregation_strategy: mean_top_k  # used to convert segment confidences into a file-level `global_confidence` score
  mean_top_k: 2               # top segments to average for global_confidence
  pred_global_threshold: 0.6  # applied to global_confidence for file-level prediction
  pred_local_threshold: 0.5   # threshold for local binary per-segments predictions

Parameter	Description
`aggregation_strategy`	How segment confidences are combined into a file-level `global_confidence` score. `"mean_top_k"` averages the top K most confident segments; `"mean_thresholded"` averages only segments exceeding `pred_local_threshold`.
`mean_top_k`	Number of top segments to average when using `mean_top_k` strategy.
`pred_global_threshold`	Threshold (0–1) applied to the aggregated global confidence to produce the final binary file-level prediction.
`pred_local_threshold`	Confidence threshold (0–1) for per-segment binary predictions (used to diplay in moderator UI). Also selects which segments contribute to global confidence under `mean_thresholded`.

Refer to repository above for complete setup and configuration details.

Technical Specifications

Model Architecture

Single-channel input → (1, 256, 312) channel x frequency x time mel spectrogram
  ↓
ResNet50 backbone (3,4,6,3 Bottleneck blocks) → (2048, 8, 10) feature map
  ↓
AdaptiveConcatPool2d [max, avg] → 4096 features
  ↓
BatchNorm1d(4096) → Dropout(0.25) → Linear(512) → ReLU
  ↓
BatchNorm1d(512) → Dropout(0.5) → Linear(2)
  ↓
Softmax → [P(negative), P(positive)]

Training Data

Source: Pod.Cast data archive from the Orcasound hydrophone network (Puget Sound, WA)
Positive examples: Confirmed annotated SRKW call segments (~0.5-4.0 seconds)
Negative examples: Background ocean noise, boats, other marine animal sounds from "in the wild" deployment
Preprocessing: Audio → Mel spectrogram (20kHz audio, 256 mel filterbank)

Training Hyperparameters

Architecture: ResNet50 (3,4,6,3 Bottleneck blocks)
Pooling: AdaptiveConcatPool2d (concatenates max + average)
Head: BN(4096) → Dropout(0.25) → Linear(512) → ReLU → BN(512) → Dropout(0.5) → Linear(2)
Data loading: SpecAugment-style augmentation with frequency masking, annotated calls padded/cropped to fixed 4.0s windows (312 time frames)
Loss: Cross-entropy
Framework: FastAI (original training)

Citation

BibTeX:

Example below:

bibtex

@misc{akash_mahajan_2026,
	author       = { Akash Mahajan and Prakruti Gogia and Aayush Agrawal },
	title        = { orcahello-srkw-detector-v1 (Revision 6ccff28) },
	year         = { 2020 },
	url          = { https://huggingface.co/orcasound/orcahello-srkw-detector-v1 },
	doi          = { 10.57967/hf/7703 },
	publisher    = { Hugging Face },
}

Grab the latest DOI with revision from the model card here.

License

This model is released under the OrcaHello RAIL (Responsible AI License), which includes specific restrictions to promote conservation of endangered Southern Resident Killer Whales.

Key restrictions:

Prohibits use in violation of Marine Mammal Protection Act
Prohibits support for captive whale industry
Requires adherence to "Be Whale Wise" guidelines

Refer to the included LICENSE file for complete terms.

Notes

Environmental Impact

This model supports conservation of the critically endangered Southern Resident Killer Whale population (currently ~75 individuals as of Jan 2026). As a component of OrcaHello’s live monitoring pipeline, it helps:

Filter 24hr hydrophone audio down to likely SRKW-call candidates for review
Enable human-in-the-loop confirmation by experts before sending alerts/notifications
Support downstream mitigation actions (e.g., coordinated vessel slow-downs and pausing pile-driving) during confirmed whale presence
Engage citizen scientists in conservation via notifications on the Orcasound live listening web-app

Learn more: https://ai4orcas.net/orcahello/

Recommendations

Calibrate confidence thresholds based on your specific deployment environment
Use as part of a two-stage detection system with expert review (moderator validation) for reliable alerts
Ensemble with fine-grained classification/captioning/analysis models for efficient processing of audio archives

Bias, Risks, and Limitations

Training data bias: Model trained primarily on Orcasound Lab hydrophone data from Puget Sound
Environmental specificity: Performance may vary with different acoustic environments
False positives: Background noise, boats, and other marine mammals may trigger false detections
Time sensitivity: Model was trained to process fixed-length 4-second segments, shorter segments padded with zeros

Contact

Project: https://www.orcasound.net/
Repository: https://github.com/orcasound/aifororcas-livesystem
Issues: https://github.com/orcasound/aifororcas-livesystem/issues