bioamla 0.0.49

Bioacoustics & Machine Learning Applications

Bioacoustics and Machine Learning Applications (bioamla)

A Python package for audio analysis and machine learning-based audio classification, focusing on bioacoustic data. Bioamla specializes in wildlife sound analysis using Audio Spectrogram Transformer (AST) models.

Prerelease Notice: This is a prerelease version of bioamla. The package is functional and ready for use, but additional features, improvements, and documentation updates are planned for 2026.

Description

Bioamla provides a toolkit for researchers, biologists, and machine learning engineers working with environmental sound data and species identification. The package combines robust audio processing capabilities with deep learning models to enable:

• Audio Classification: Classify wildlife sounds, species calls, and environmental audio using pre-trained or fine-tuned AST models
• Model Training: Fine-tune Audio Spectrogram Transformer models on custom datasets from Hugging Face Hub
• Batch Processing: Efficiently process directories of audio files with temporal segmentation
• Audio Processing: Load, resample, split, and extract metadata from various audio formats
• System Diagnostics: Monitor GPU/CUDA availability and package versions

Setup

Requirements

• Python 3.8 or higher
• CUDA-capable GPU (recommended for training and large-scale inference)

Installation w/ pip

Install bioamla using pip:

pip install bioamla

Installation from Source

For development or the latest version:

git clone https://github.com/jmcmeen/bioamla.git
cd bioamla
pip install -e .

Verify Installation

Check that bioamla is installed correctly:

bioamla version
bioamla devices

The devices command will show your CUDA/GPU availability and configuration.

Examples

1. Basic Audio Classification

Classify a single audio file using a pre-trained model:

bioamla ast predict path/to/audio.wav --model-path bioamla/scp-frogs

This will output the predicted class for the audio file.

2. Batch Inference on Directory

Process all audio files in a directory and export results to CSV using the --batch flag:

bioamla ast predict /path/to/audio/directory --batch \
  --output-csv results.csv \
  --model-path bioamla/scp-frogs \
  --resample-freq 16000 \
  --clip-seconds 1 \
  --overlap-seconds 0.5

This creates a CSV file with columns: filepath, start, stop, prediction

Optimized inference with GPU acceleration:

bioamla ast predict /path/to/audio/directory --batch \
  --model-path bioamla/scp-frogs \
  --batch-size 16 \
  --fp16 \
  --compile \
  --workers 4

Performance options (batch mode only):

• --batch-size: Process multiple segments in one forward pass (2-4x faster)
• --fp16: Use half-precision inference (~2x faster on modern GPUs)
• --compile: Use torch.compile() for optimized execution (1.5-2x faster, PyTorch 2.0+)
• --workers: Parallel file loading for I/O-bound workloads

Resume interrupted processing:

bioamla ast predict /path/to/audio/directory --batch \
  --output-csv results.csv \
  --no-restart

3. Fine-tune a Model

Train a custom model on your own dataset from Hugging Face Hub:

bioamla ast train \
  --training-dir ./my-training \
  --base-model MIT/ast-finetuned-audioset-10-10-0.4593 \
  --train-dataset your-username/your-dataset \
  --num-train-epochs 10 \
  --learning-rate 5e-5 \
  --per-device-train-batch-size 8 \
  --eval-strategy epoch \
  --report-to tensorboard

Monitor training with TensorBoard:

tensorboard --logdir ./my-training

Push trained model to Hugging Face Hub:

bioamla ast train \
  --training-dir ./my-training \
  --train-dataset your-username/your-dataset \
  --num-train-epochs 10 \
  --push-to-hub

4. Model Evaluation

Evaluate model performance on a test dataset with ground truth labels:

bioamla ast evaluate ./test_audio --model-path bioamla/scp-frogs \
  --ground-truth labels.csv

This outputs:

• Accuracy, Precision, Recall, F1 Score
• Per-class metrics
• Confusion matrix

Save results to different formats:

# JSON output
bioamla ast evaluate ./test_audio --ground-truth labels.csv \
  --output results.json --format json

# CSV output (per-class metrics)
bioamla ast evaluate ./test_audio --ground-truth labels.csv \
  --output results.csv --format csv

5. Spectrogram Visualization

Generate spectrograms and other audio visualizations:

Single file:

bioamla visualize audio.wav --output spectrogram.png

Batch processing:

bioamla visualize ./audio_dir --batch --output ./spectrograms

Visualization types:

# Mel spectrogram (default)
bioamla visualize audio.wav --type mel --output mel_spec.png

# MFCC visualization
bioamla visualize audio.wav --type mfcc --output mfcc.png

# Waveform plot
bioamla visualize audio.wav --type waveform --output waveform.png

6. Audio Augmentation

Expand training datasets with augmented audio:

bioamla augment ./audio --output ./augmented \
  --add-noise 3-30 \
  --time-stretch 0.8-1.2 \
  --pitch-shift -2,2 \
  --multiply 5

This creates 5 augmented copies of each audio file with random combinations of:

• Gaussian noise (SNR 3-30 dB)
• Time stretching (80%-120% speed)
• Pitch shifting (-2 to +2 semitones)

7. Signal Processing

Process audio files with filtering, denoising, and other operations:

Apply frequency filters:

# Bandpass filter (keep 1000-8000 Hz)
bioamla audio filter recording.wav --bandpass 1000-8000 --output filtered.wav

# Lowpass filter (remove high frequencies)
bioamla audio filter recording.wav --lowpass 4000 --output lowpassed.wav

# Highpass filter (remove low frequencies)
bioamla audio filter recording.wav --highpass 500 --output highpassed.wav

Remove noise:

bioamla audio denoise noisy.wav --output clean.wav --strength 1.5

Split audio on silence:

bioamla audio segment long_recording.wav --output ./segments \
  --silence-threshold -40 \
  --min-silence 0.3 \
  --min-segment 0.5

Detect onset events:

bioamla audio detect-events recording.wav --output events.csv

Normalize loudness:

# RMS normalization (default)
bioamla audio normalize recording.wav --target-db -20 --output normalized.wav

# Peak normalization
bioamla audio normalize recording.wav --peak --target-db -3 --output normalized.wav

Resample audio:

bioamla audio resample recording.wav --rate 16000 --output resampled.wav

Trim audio:

# Trim by time
bioamla audio trim recording.wav --start 1.5 --end 5.0 --output trimmed.wav

# Trim silence from start and end
bioamla audio trim recording.wav --silence --threshold -40 --output trimmed.wav

Batch processing:

All signal processing commands support --batch for directory processing:

bioamla audio normalize ./recordings --batch --output ./normalized --target-db -20
bioamla audio filter ./recordings --batch --output ./filtered --lowpass 8000

8. Audio File Utilities

List all audio files in a directory:

bioamla audio list /path/to/audio/directory

Extract WAV file metadata:

bioamla audio info /path/to/file.wav

Download audio files:

bioamla download https://example.com/audio.zip ./downloads

Extract archives:

bioamla unzip ./downloads/audio.zip ./extracted

9. License File Generation

Generate license/attribution files from dataset metadata:

Single dataset:

bioamla dataset license ./my_dataset

With a template file:

bioamla dataset license ./my_dataset --template ./license_template.txt

Process all datasets in a directory:

bioamla dataset license ./audio_datasets --batch

Custom output filename:

bioamla dataset license ./my_dataset --output ATTRIBUTION.txt

The metadata CSV must contain these columns: file_name, attr_id, attr_lic, attr_url, attr_note

10. Python API Usage

Use bioamla programmatically in your Python scripts:

from bioamla.ast import load_pretrained_ast_model, wav_ast_inference
from bioamla.torchaudio import load_waveform_tensor

# Load a pre-trained model
model, processor = load_pretrained_ast_model("bioamla/scp-frogs")

# Run inference on a single file
predictions = wav_ast_inference(
    wav_filepath="path/to/audio.wav",
    model_path="bioamla/scp-frogs",
    resample_freq=16000
)

# Print top predictions
for pred in predictions[:5]:
    print(f"{pred['label']}: {pred['score']:.4f}")

Batch processing with segmentation:

from bioamla.ast import wave_file_batch_inference

# Process directory with 1-second segments and 0.5s overlap
wave_file_batch_inference(
    directory="./audio_files",
    model_path="bioamla/scp-frogs",
    output_csv="results.csv",
    resample_freq=16000,
    clip_seconds=1,
    overlap_seconds=0.5,
    restart=True
)

Load and process audio:

from bioamla.torchaudio import (
    load_waveform_tensor,
    resample_waveform_tensor,
    split_waveform_tensor
)

# Load audio file
waveform, sample_rate = load_waveform_tensor("audio.wav")

# Resample to 16kHz
waveform_resampled = resample_waveform_tensor(
    waveform, sample_rate, 16000
)

# Split into 1-second segments with 0.5s overlap
segments = split_waveform_tensor(
    waveform_resampled,
    sample_rate=16000,
    clip_seconds=1,
    overlap_seconds=0.5
)

11. System Diagnostics

Check GPU availability:

from bioamla.diagnostics import get_device_info

device_info = get_device_info()
print(f"CUDA available: {device_info['cuda_available']}")
print(f"Device count: {device_info['device_count']}")
print(f"Device name: {device_info['device_name']}")

Get package versions:

from bioamla.diagnostics import get_package_versions

versions = get_package_versions()
for package, version in versions.items():
    print(f"{package}: {version}")

12. Dataset Explorer (Experimental)

Launch an interactive terminal dashboard to explore audio datasets:

bioamla explore ./my_dataset

The explorer provides:

• File browser with sorting and filtering
• Dataset statistics (total files, size, formats)
• Category and split summaries (if metadata.csv present)
• Audio playback (requires system audio player)
• Spectrogram generation and viewing
• Search functionality

13. Experiment Tracking with MLflow

bioamla integrates with MLflow for experiment tracking during model training:

Start MLflow server:

mlflow server --host 0.0.0.0 --port 5000

Train with MLflow tracking:

bioamla ast train \
  --training-dir "my-model" \
  --train-dataset "bioamla/scp-frogs" \
  --num-train-epochs 10 \
  --mlflow-tracking-uri "http://localhost:5000" \
  --mlflow-experiment-name "frog-classifier" \
  --mlflow-run-name "baseline-run"

View experiments in MLflow UI:

Open http://localhost:5000 in your browser to view training metrics, compare runs, and analyze model performance.

MLflow tracks:

• Training and evaluation metrics (loss, accuracy)
• Model hyperparameters
• Training artifacts

CLI Reference

Command	Description
bioamla version	Display bioamla version
bioamla devices	Show CUDA/GPU information
bioamla explore <DIR>	Launch interactive TUI dashboard for exploring datasets
bioamla purge	Purge cached HuggingFace Hub data (models/datasets)
bioamla visualize <PATH>	Generate spectrogram visualizations
bioamla augment <INPUT_DIR>	Augment audio files to expand training datasets
bioamla download <URL> [DIR]	Download files from URL
bioamla unzip <FILE> [DIR]	Extract ZIP archives
bioamla zip <SOURCE> <OUTPUT>	Create ZIP archive from file or directory

Audio Commands (bioamla audio)

Command	Description
bioamla audio list [DIR]	List audio files in directory
bioamla audio info <FILE>	Display WAV file metadata
bioamla audio convert <PATH> <FORMAT>	Convert audio files between formats
bioamla audio filter <PATH>	Apply frequency filters (bandpass, lowpass, highpass)
bioamla audio denoise <PATH>	Apply spectral noise reduction
bioamla audio segment <PATH>	Split audio on silence into separate files
bioamla audio detect-events <PATH>	Detect onset events and export to CSV
bioamla audio normalize <PATH>	Normalize audio loudness (RMS or peak)
bioamla audio resample <PATH>	Resample audio to a different sample rate
bioamla audio trim <PATH>	Trim audio by time or remove silence

AST Model Commands (bioamla ast)

Command	Description
bioamla ast predict <PATH>	Single file or batch inference
bioamla ast train	Fine-tune AST model on custom datasets
bioamla ast evaluate <PATH>	Evaluate model on test data with ground truth labels

iNaturalist Commands (bioamla inat)

Command	Description
bioamla inat download <OUTPUT_DIR>	Download audio observations from iNaturalist
bioamla inat search	Search for taxa with observations in a place or project
bioamla inat stats <PROJECT_ID>	Get statistics for an iNaturalist project

Dataset Commands (bioamla dataset)

Command	Description
bioamla dataset merge <OUTPUT_DIR> <PATHS...>	Merge multiple audio datasets into one
bioamla dataset license <PATH>	Generate license/attribution file from metadata

HuggingFace Hub Commands (bioamla hf)

Command	Description
bioamla hf push-model <PATH> <REPO_ID>	Push model folder to HuggingFace Hub
bioamla hf push-dataset <PATH> <REPO_ID>	Push dataset folder to HuggingFace Hub

Both push commands automatically detect large folders (>5GB or >1000 files) and use the optimized upload_large_folder method for better reliability.

Use bioamla <command> --help for detailed options on any command.

Technologies

• PyTorch + HuggingFace Transformers: Audio Spectrogram Transformer models
• TorchAudio: Audio file I/O and preprocessing
• Librosa: Audio analysis and feature extraction
• SciPy: Signal processing and filtering
• Click: Command-line interface framework
• Textual: Terminal user interface for dataset exploration
• FastAPI: Web service capability (optional)
• Pydantic: Data validation and API schemas
• Audiomentations: Audio data augmentation
• Matplotlib: Spectrogram visualization
• TensorBoard: Training visualization
• MLflow: Experiment tracking and model management (optional)

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Citation

If you use bioamla in your research, please cite:

@software{bioamla,
  author = {McMeen, John},
  title = {Bioamla: Bioacoustics and Machine Learning Applications},
  year = {2025},
  url = {https://github.com/jmcmeen/bioamla}
}