Bird Sound Classifier using Custom CNN

This repository contains code to build, train, and use a custom Convolutional Neural Network (CNN) for identifying bird species from audio recordings. The system processes recordings from the Xeno-Canto database, trains on spectrograms of bird sounds, and can analyze new audio files to classify species.

---

Features

• Fetch bird sound recordings from the Xeno-Canto database.
• Preprocess audio files into spectrograms suitable for CNN input.
• Train a CNN model to classify bird sounds.
• Use the trained model to analyze query audio files for bird species identification.
• Generate detailed outputs including classifications, timestamps, and visualizations.

---

Installation

1. Clone this repository:

   git clone https://github.com/your-username/bird-sound-classifier.git
   cd bird-sound-classifier

2. Install the required R packages:

   install.packages(c("tidyverse", "warbleR", "av", "tuneR", "snow", "furrr", 
                      "fs", "abind", "caret", "keras", "e1071", "pheatmap", "RColorBrewer"))

3. Install TensorFlow for Keras:

   library(keras)
   install_keras()

---

Usage

1. Training the Model

This part involves preprocessing audio data, building a CNN model, and training it.

Preprocessing:

• Fetch and balance bird sound recordings for target and background species.
• Convert audio files into spectrograms and stratify data into training, validation, and testing sets.

Training:

• Use the following code to train the CNN:

  source("funs.R") # Load utility functions
  
  # Train the model
  history <- fit(model, x = train$X, y = train$Y,
                 batch_size = 16, epochs = 15,
                 validation_data = list(val$X, val$Y))

Evaluation:

• Validate the model and test its accuracy on unseen data.

2. Using the Model

Once trained, the model can be used to classify bird species from a query audio file.

Step-by-step Instructions:

1. Prepare the Query Audio File:

   • Split the audio into overlapping windows of specified size and stride.

   query = "path/to/your/query_audio.mp3"
   windsiz = 10 # Window size in seconds
   strid = 5    # Stride length in seconds
   queryX <- audioProcess(files = query, limit = (query_dur - windsiz), ws = windsiz, stride = strid)

2. Load the Model and Predict:

   • Load the trained CNN and classify species in each time window.

   model <- load_model_tf(filepath = "path/to/your/saved_model")
   predXquery <- predict(model, queryX)

3. Output Results:

   • Create a detailed table of predictions, including timestamps, species IDs, and classification accuracy.
   • Filter results by accuracy and visualize them using a pie chart and spectrogram.

   # Create summary table
   queryTable <- data.frame(cbind(timestamp, predXClass, as.numeric(accuracy)))
   queryTable <- filter(queryTable, accuracy >= 0.9 & ID != "no class")
   
   # Generate pie chart
   ggplot(queryTable, aes(x="", y=n, fill=ID)) +
     geom_bar(stat="identity", width=1, color="white") +
     coord_polar("y", start=0) +
     labs(title = "Pie chart of detected species") +
     theme_void()
   
   # Spectrogram visualization
   image(queryX[33,,,],
         xlab = "Time (s)",
         ylab = "Frequency (kHz)",
         axes = F)

---

Example Workflow

1. Train the Model:

   • Preprocess audio and train the CNN as described above.

2. Test a Query Audio File:

   • Use the code provided in the "Using the Model" section to predict bird species in a new audio file.

3. Visualize Results:

   • Generate visualizations like pie charts of detected species and spectrograms of audio segments.

---

Dependencies

• Xeno-Canto API
• R (≥ 4.0.0)
• Keras with TensorFlow backend

---

Acknowledgments

• Bird sound data sourced from the Xeno-Canto database.
• The warbleR package for Xeno-Canto queries and audio handling.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.