Analysis of individual identification and age-class classification of wild female macaque vocalizations without pitch- and formant-based acoustic parameter measurements

This study demonstrates that deep learning-derived mel spectrograms, when combined with traditional machine learning classifiers like random forests and support vector machines, effectively enable individual identification and age-class classification of wild female Japanese macaque vocalizations even with small-scale datasets.

The Gist

Imagine you are at a massive, noisy concert where everyone is wearing identical masks. You can't see their faces, but you can hear them singing. Your goal? To figure out exactly who is singing a specific note, and also to guess whether that singer is a teenager or a grandparent, just by listening.

This is exactly what the researchers in this paper did, but instead of a concert, they were listening to wild Japanese macaques (monkeys) on Yakushima Island.

Here is the story of their experiment, broken down into simple concepts:

1. The Problem: The "Small Data" Dilemma

In the world of Artificial Intelligence (AI), computers usually need to "eat" thousands or millions of examples to learn how to recognize things. It's like trying to teach a child to recognize a cat by showing them a million photos of cats.

But in the wild, getting that many recordings is nearly impossible. You can't just ask a monkey to sing 1,000 times for you! The researchers had a small pile of data: 651 "coo" calls (a friendly contact call) from just six female monkeys. They wanted to know: Can modern AI learn to identify these specific monkeys and guess their age with such a tiny dataset, without us manually measuring every single sound wave?

2. The Tool: The "Mel Spectrogram" (The Sound Fingerprint)

Traditionally, scientists would act like audio engineers, manually measuring specific things like "how high the pitch is" or "how long the call lasts." This is like trying to describe a painting by only listing the exact shade of blue and the size of the brushstrokes. It's tedious and you might miss the big picture.

Instead, this team used a Mel Spectrogram.

• The Analogy: Imagine taking a photo of the sound. Instead of just measuring the height of the wave, you turn the sound into a colorful, visual map (like a heat map) that shows how the sound changes over time.
• Why it works: It captures the "vibe" of the sound—the texture, the roughness, the flow—without needing a human to pick out specific numbers. It's like letting the AI look at the whole painting rather than just counting the pixels.

3. The Challenge: Two Missions

The AI had to pass two tests:

1. Mission A (The Detective): "Who sang this?" (Identify the specific monkey).
2. Mission B (The Age Guess): "Is this a young monkey (under 10) or an old monkey (over 20)?"

4. The Results: A Surprise Success!

The researchers used two different types of AI "brains" (Random Forest and Support Vector Machines) to analyze the sound maps.

• Mission A (Who is it?): The AI got it right about 81-82% of the time.

  • The Catch: It was great at spotting some monkeys (like "Sasa") but got confused between a few others (like "Kapa" and "Rine"). It's like having a friend who can instantly recognize your best friend's laugh but sometimes mixes up your two cousins who sound similar.
  • Why it matters: This is a huge win because they did this with wild, noisy recordings, not a perfect studio setup.

• Mission B (How old?): The AI crushed this test, getting it right 91-93% of the time!

  • The Secret: The older monkeys' voices had a specific "texture" (perhaps a bit rougher or "harsher," like an old vinyl record compared to a crisp CD). The AI picked up on this subtle "aging" texture automatically, even though the researchers didn't tell it to look for "roughness."

5. The Big Picture: Why This Matters

Think of this research as building a non-invasive ID badge system for nature.

• No more tagging: Usually, to count animals or study them, scientists have to trap them and put a collar or a tag on them. This is stressful for the animal.
• The New Way: With this method, scientists can just set up a microphone in the forest. The AI listens, says, "That's Monkey #4, and she's getting on in years," and logs it.
• The "Small Data" Breakthrough: The most important takeaway is that you don't need a million samples to make this work. If you have a decent amount of data (a few hundred calls), modern AI can learn the "fingerprint" of a wild animal just fine.

In a Nutshell

The researchers proved that you don't need to be a sound engineer to teach a computer to recognize wild monkeys. By letting the computer look at the "sound picture" (the spectrogram) rather than measuring specific notes, they successfully built a system that can identify individual monkeys and guess their age, even with a small amount of data. It's a giant leap toward letting technology help us understand nature without disturbing it.

Technical Summary

1. Problem Statement

Bioacoustic studies have traditionally relied on manually measuring specific acoustic features (e.g., pitch, formants, duration) to classify animal vocalizations for individual identification or repertoire analysis. However, these conventional methods suffer from:

• Arbitrariness: The selection of features is often subjective and species-specific, limiting versatility.
• Data Scarcity: Deep learning models, which can automatically process acoustic signals without manual feature extraction, typically require massive datasets (thousands to millions of samples). This is a significant barrier in wildlife research where collecting large amounts of labeled field data is time-consuming and often impossible.
• Gap: There is a lack of validated methods for classifying relatively small-scale wild acoustic datasets using automatically processed features (like mel spectrograms) without explicit manual feature engineering.

2. Methodology

Data Collection

• Subjects: Wild Japanese macaques (Macaca fuscata yakui) on Yakushima Island, Japan.
• Focal Group: The "Petit" group (approx. 33 individuals).
• Sample: Vocalizations from 6 adult females were selected for analysis based on data quality and quantity.
  • Younger group (<10 yrs): Kapa (9), Rine (8), Sasa (7), Taiwu (9).
  • Older group (>20 yrs): Sazae (21-22), Taiko (23-25).
  • Note: Two pairs were mother-daughter (Sazae/Sasa and Taiko/Taiwu).
• Recording: 112 hours of field recording (Oct–Dec 2023) using directional and lavalier microphones.
• Selection Criteria: 651 "coo" calls (contact calls) were selected after strict filtering: low background noise, no overlap, confirmed vocalizer, and duration <1 second.

Feature Extraction & Preprocessing

• Feature Type: Mel Spectrograms (automatically processed acoustic features representing the short-term power spectrum of sound on a non-linear mel scale).
• Parameters: Generated using Python's librosa with parameters based on previous meerkat studies:
  • 40 Mel bands.
  • 30ms window size, 3.75ms hop size.
  • Frequency range: 0–24,000 Hz.
  • Zero-padding applied to match the longest call duration.
  • Z-transformation applied to normalize intensity.
• Dimensionality: Each call resulted in a 10,360-dimensional vector.

Analysis & Classification Tasks

Two primary tasks were evaluated:

1. Individual Identification: Distinguishing between the 6 specific females.
2. Age-Class Classification: Distinguishing between "Younger" (<10 yrs) and "Older" (>20 yrs).

Algorithms:

• Random Forest (RF) and Support Vector Machine (SVM).
• Validation Strategy: To address small sample sizes and class imbalance, the authors performed 1,000 iterations of train-test splits. In each iteration, 80 calls per individual were randomly selected (64 for training, 16 for testing).
• Hyperparameter Tuning: Grid search with 5-repeated 5-fold cross-validation was used to find optimal parameters for RF and SVM.
• Evaluation Metric: Balanced Accuracy (average of recall per class) to mitigate bias from unequal sample sizes.

Dimensionality Reduction:

• UMAP (Uniform Manifold Approximation and Projection): Used to visualize data structure. Both Supervised UMAP (using labels to guide projection) and Unsupervised UMAP (label-free) were compared using silhouette scores and inter/intra-class distances.

3. Key Results

Individual Identification

• Performance:
  • Random Forest: Mean Balanced Accuracy = 0.81.
  • SVM: Mean Balanced Accuracy = 0.82.
• Variability: Accuracy varied by individual (Range: 0.68–0.94). Sasa (youngest) had the highest accuracy; Kapa had the lowest.
• Error Analysis: Misidentifications occurred primarily among Kapa, Rine, and Taiwu.
• Data Distribution: Calls were bimodally distributed into "high accuracy" (>0.9) and "low accuracy" (<0.16). 447 calls were high accuracy; 55 were low accuracy.
• UMAP Findings: Supervised UMAP showed clear clustering by individual (Mean Silhouette Score: 0.60), whereas Unsupervised UMAP showed poor separation (0.05).

Age-Class Classification

• Performance:
  • Random Forest: Mean Balanced Accuracy = 0.91.
  • SVM: Mean Balanced Accuracy = 0.93.
• Class-Specific Accuracy:
  • **Younger (<10 yrs):** >98% accuracy.
  • Older (>20 yrs): <87% accuracy.
• Error Analysis: Errors were almost exclusively misclassifying older calls as younger. Only 23 calls (from Sazae and Taiko) were consistently low-accuracy across both models.
• UMAP Findings: Supervised UMAP showed tight clustering by age class (Silhouette scores: 0.86 for young, 0.80 for old), significantly outperforming unsupervised methods.

4. Key Contributions

1. Validation of Small-Scale Deep Learning Features: Demonstrated that mel spectrograms (typically associated with deep learning) are effective for classification tasks even with small datasets (~100 calls per class) when paired with traditional machine learning classifiers (RF/SVM).
2. Elimination of Manual Feature Engineering: Proved that high-accuracy identification and age classification can be achieved without measuring specific acoustic parameters like pitch or formants, reducing researcher bias and species-specific constraints.
3. Age Classification in Wild Primates: Successfully classified age classes in wild macaques, achieving >90% accuracy, suggesting that mel spectrograms capture subtle, complex age-related acoustic changes (e.g., "harshness") that manual metrics might miss.
4. Methodological Framework: Provided a robust pipeline (UMAP visualization + repeated train-test splits) for evaluating bioacoustic models on limited field data.

5. Significance and Discussion

• Wildlife Monitoring: The high accuracy (80%+) in individual identification suggests this method is viable for non-invasive mark-recapture studies and population density estimation in the wild, including nocturnal monitoring.
• Senescence Research: The study highlights that vocal aging in macaques produces distinct acoustic signatures detectable by automated systems. This opens new avenues for studying senescence in wild populations without capturing animals.
• Limitations:
  • Sample Size: Only two older individuals were used, meaning the "age" classifier might be learning specific individual traits of Sazae and Taiko rather than general aging features.
  • Generalizability: Further research is needed to test if these results hold for larger groups, males, and different age ranges.
  • Interpretability: While the models work, the specific acoustic properties driving the age classification (e.g., specific spectral noise patterns) remain implicit and require further investigation.

Conclusion: The paper establishes that automatically processed acoustic features (mel spectrograms) combined with standard machine learning classifiers are a powerful, efficient tool for analyzing small-scale wild primate vocal datasets, offering a scalable alternative to traditional manual acoustic analysis.