BirdNET Batch Analysis for Audacity Integration

---

Abstract

This Google Colab notebook provides a batch audio processing pipeline for bioacoustic analysis using BirdNET (Cornell Lab of Ornithology). The workflow merges multiple audio recordings end-to-end, runs BirdNET species detection with confidence thresholds, generates Audacity-compatible label files for visual inspection, and produces comprehensive visualizations including waveforms, spectrograms, f0 tracks, and formant-like resonance estimates (F1-F3) via Linear Predictive Coding (LPC).

Key Features:

• End-to-end audio merging with timeline index generation
• BirdNET automated bird call detection (7-class CNN)
• Audacity label export for manual verification
• Per-detection waveform and spectrogram visualization
• F0 estimation via librosa pyin (YIN algorithm)
• Formant estimation via LPC autocorrelation method
• Overview spectrogram with detection overlays

Application: Ornithology research, biodiversity monitoring, acoustic ecology, bioacoustic annotation workflows.

---

1. Introduction

1.1 BirdNET

BirdNET (Kahl et al., 2021) is a deep learning model for automated bird species identification from audio recordings. Developed by the Cornell Lab of Ornithology and Chemnitz University of Technology, it uses a convolutional neural network (CNN) trained on millions of bird vocalizations to detect and classify over 3,000 species globally.

Architecture:

• Input: Mel spectrogram (time-frequency representation)
• CNN backbone: Residual blocks with batch normalization
• Output: Softmax classification with confidence scores (0-1)

1.2 Workflow Overview

The pipeline consists of four stages:

1. Upload & Merge: Concatenate audio files end-to-end → export merged WAV + Audacity labels + index CSV
2. BirdNET Analysis: Run detection on original files → align detections to merged timeline
3. Visualize Detections: Generate waveforms + spectrograms with f0/formant overlays
4. Overview Spectrogram: Full-file spectrogram with detection spans labeled

---

2. Mathematical Foundations

2.1 F0 Estimation (pyin / YIN Algorithm)

YIN (de Cheveigné & Kawahara, 2002) estimates fundamental frequency via autocorrelation:

Difference function: $$ d(\tau) = \sum_{n=0}^{N-\tau-1} (x[n] - x[n+\tau])^2 $$

Cumulative mean normalized difference: $$ d'(\tau) = \begin{cases} 1 & \tau = 0 \ \frac{d(\tau)}{\frac{1}{\tau} \sum_{j=1}^{\tau} d(j)} & \tau > 0 \end{cases} $$

Period estimation: First minimum of $d'(\tau)$ below threshold (typically 0.1).

Fundamental frequency: $$ f_0 = \frac{f_s}{\hat{\tau}}, \quad \hat{\tau} = \arg\min_{\tau} d'(\tau) $$

pyin (Mauch & Dixon, 2014) extends YIN with probabilistic voicing detection for more robust f0 tracking.

2.2 Formant Estimation (Linear Predictive Coding)

LPC models the vocal tract as an all-pole filter:

$$ H(z) = \frac{G}{1 - \sum_{k=1}^{p} a_k z^{-k}} $$

Autocorrelation method (Levinson-Durbin recursion):

$$ R(m) = \sum_{n=0}^{N-m-1} x[n] \cdot x[n+m], \quad m = 0, 1, \ldots, p $$

Solve for coefficients $a_1, \ldots, a_p$ via:

$$ \sum_{k=0}^{p} a_k R(|m-k|) = 0, \quad m = 1, \ldots, p $$

Formant frequencies from LPC roots:

$$ f_i = \frac{\text{angle}(r_i)}{2\pi} \cdot f_s $$

where $r_i$ are complex roots of $A(z) = 1 - \sum_{k=1}^{p} a_k z^{-k}$ with positive imaginary parts.

Note: For bird vocalizations, "formants" represent spectral resonances rather than true vocal tract formants (as in speech).

2.3 Mel Spectrogram

Mel scale (perceptually-motivated frequency scale):

$$ \text{Mel}(f) = 2595 \log_{10}\left(1 + \frac{f}{700}\right) $$

Mel filterbank: Triangular filters spaced linearly on Mel scale, log-linearly on Hz.

BirdNET preprocessing: STFT → Mel filterbank → log magnitude → CNN input.

---

3. Pipeline Stages

3.1 Stage 1: Upload & Merge

Process:

1. Upload audio files to Google Colab
2. Sort by filename (alphabetical) or modification time
3. Concatenate end-to-end with optional gap (default: 0 ms)
4. Normalize loudness (per-file and final mix)
5. Export merged WAV + Audacity labels + index CSV

Outputs:

• merged_no_overlap.wav — Concatenated audio
• merged_no_overlap_labels_audacity.txt — File boundaries as Audacity regions
• merged_no_overlap_index.csv — Timeline mapping (filename → start/end times)

3.2 Stage 2: BirdNET Analysis

Process:

1. Load index CSV from Stage 1
2. Run BirdNET on each original file
3. Align detections to merged timeline using start offsets
4. Filter by confidence threshold (default: 0.25)
5. Export detections CSV + Audacity labels

Outputs:

• merged_no_overlap_birdnet_detections.csv — All detections with global timestamps
• merged_no_overlap_birdnet_labels.txt — Detection regions for Audacity

Configuration:

• MIN_CONF: Minimum confidence (0-1)
• LAT, LON, DATE: Optional geographic/temporal context for species filtering
• SPECIES_FILTER: Optional list of expected species (reduces false positives)

3.3 Stage 3: Visualize Detections

Process (per detection):

1. Extract audio segment (with padding: default ±0.2s)
2. Compute f0 track via librosa pyin
3. Estimate formants (F1-F3) via LPC (order = max(8, 2 + sr/1000))
4. Calculate spectral features (centroid, peak frequency)
5. Generate waveform plot (time domain)
6. Generate spectrogram with f0 overlay (frequency domain)

Outputs:

• det_XXXX_waveform.png — Time-domain amplitude plot
• det_XXXX_spectrogram.png — STFT spectrogram with f0 track overlay
• birdnet_detection_summary.csv — Metrics table (f0, F1-F3, spectral centroid, etc.)

3.4 Stage 4: Overview Spectrogram

Process:

1. Compute full-file STFT spectrogram
2. Overlay BirdNET detections as shaded time spans
3. Label each detection with species name + confidence

Output:

• overview_spectrogram.png — Full timeline spectrogram with annotations

---

4. Project Structure

bird-net-batch-analysis/
├── AudacityTaggerBatchCompiler.ipynb   # Google Colab notebook (all stages)
└── README.md

---

5. Installation and Usage

5.1 Google Colab (Recommended)

1. Open notebook: Click "Open in Colab" badge in AudacityTaggerBatchCompiler.ipynb
2. Run setup cells: Installs dependencies (BirdNET, librosa, pydub, graphviz)
3. Upload audio files: Use file upload widget in Cell 1
4. Configure parameters:
   • Stage 1: NORMALIZE_EACH, GAP_BETWEEN_MS, FORCE_SAMPLE_RATE
   • Stage 2: MIN_CONF, LAT, LON, DATE
   • Stage 3: F0_MIN_HZ, F0_MAX_HZ, LPC_ORDER, PAD_BEFORE_S, PAD_AFTER_S
5. Run cells sequentially: Execute Shift+Enter for each cell
6. Download outputs: All files saved to /content/ (accessible via Colab file browser)

5.2 Key Parameters

Parameter	Default	Description
MIN_CONF	0.25	Minimum BirdNET confidence (0-1)
F0_MIN_HZ	100.0	Lower bound for f0 search (Hz)
F0_MAX_HZ	8000.0	Upper bound for f0 search (Hz)
LPC_ORDER	auto	LPC order for formant estimation (typically 8-16)
PAD_BEFORE_S	0.2	Context before detection (seconds)
PAD_AFTER_S	0.2	Context after detection (seconds)
GAP_BETWEEN_MS	0	Silence between merged files (ms)

5.3 Audacity Integration

Import merged audio:

1. File → Open → Select merged_no_overlap.wav

Import labels:

1. File → Import → Labels → Select merged_no_overlap_labels_audacity.txt (file boundaries)
2. File → Import → Labels → Select merged_no_overlap_birdnet_labels.txt (detections)

Switch to spectrogram view:

1. Click track dropdown → Spectrogram
2. Navigate detections using label regions
3. Verify BirdNET classifications visually

---

6. Outputs Summary

Generated Files

File	Description
merged_no_overlap.wav	Concatenated audio (all files end-to-end)
merged_no_overlap_labels_audacity.txt	File boundary labels for Audacity
merged_no_overlap_index.csv	Timeline index (filename → start/end)
merged_no_overlap_birdnet_detections.csv	BirdNET detections with global timestamps
merged_no_overlap_birdnet_labels.txt	Detection labels for Audacity
birdnet_detection_summary.csv	Per-detection metrics (f0, F1-F3, etc.)
det_XXXX_waveform.png	Waveform plot for detection XXXX
det_XXXX_spectrogram.png	Spectrogram plot for detection XXXX
overview_spectrogram.png	Full-file spectrogram with overlays

---

7. References

BirdNET

• Kahl, S., Wood, C. M., Eibl, M., & Klinck, H. (2021). "BirdNET: A deep learning solution for avian diversity monitoring". Ecological Informatics, 61, 101236. DOI: 10.1016/j.ecoinf.2021.101236
• BirdNET GitHub: https://github.com/kahst/BirdNET-Analyzer

F0 Estimation

• de Cheveigné, A., & Kawahara, H. (2002). "YIN, a fundamental frequency estimator for speech and music". Journal of the Acoustical Society of America, 111(4), 1917-1930. DOI: 10.1121/1.1458024
• Mauch, M., & Dixon, S. (2014). "pYIN: A fundamental frequency estimator using probabilistic threshold distributions". Proc. ICASSP, 659-663.

Linear Predictive Coding

• Makhoul, J. (1975). "Linear prediction: A tutorial review". Proceedings of the IEEE, 63(4), 561-580.
• Rabiner, L. R., & Schafer, R. W. (2011). Theory and Applications of Digital Speech Processing. Prentice Hall.

Librosa

• McFee, B., et al. (2015). "librosa: Audio and Music Signal Analysis in Python". Proc. SciPy, 18-24. DOI: 10.25080/Majora-7b98e3ed-003

---

8. License

MIT License (2025)

Copyright (c) 2025 George Redpath

---

9. Author

George Redpath (Ziforge) GitHub: @Ziforge Focus: Bioacoustics, ornithology, automated species detection

---

10. Acknowledgments

• Cornell Lab of Ornithology — BirdNET model development
• Google Colab — Cloud computing platform
• Audacity Team — Open-source audio editor
• librosa Contributors — Audio analysis library

---

11. Citation

@misc{redpath2025birdnet,
  author = {Redpath, George},
  title = {BirdNET Batch Analysis for Audacity Integration},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/Ziforge/bird-net-batch-analysis}
}

---

Built for bioacoustic research and ornithological field studies.