In this project, we designed a system that takes a spoken sentence and segments it into words 🗣️✂️.
The system detects when each word starts and ends, without knowing in advance how many words there are — only assuming a small silence gap between words.

Additionally, we created a program to play each detected word separately.
Finally, the system estimates the average pitch (fundamental frequency) of the speaker.

The following classifiers were trained and evaluated:

• Least Squares (LSQ)
• Support Vector Machine (SVM)
• Multilayer Perceptron (MLP) (Three-layer neural network)
• Recurrent Neural Network (RNN)

• Programming Language: Python 3.12.4 🐍
• ❗ No CNNs, no web services, no transfer learning allowed.
• Deliverables: PDF documentation, source code (source2023.zip), auxiliary files (auxiliary2023.zip).

The system does binary classification:
✅ Speech (foreground) vs ❌ Non-speech (background).

Main Steps:

1. Extract Mel spectrograms 🎶 from sliding windows of the audio.
2. Classify each window as speech or non-speech.
3. Apply a median filter to smooth out small errors 🧹.
4. Find the boundaries between words based on the cleaned-up predictions 🧩.

• Simple models that output a continuous value.
• Then we threshold them to get binary speech/non-speech predictions.

• Two hidden layers: 128 and 64 neurons with ReLU activation ⚡.
• Output layer: Single neuron with Sigmoid activation 🧠.
• Trained with Binary Cross-Entropy Loss.

• Built using TensorFlow's SimpleRNN layers 🔁.
• Processes sequences of frames to capture temporal dynamics.
• Outputs one probability per time frame.

Foreground (Speech) 🗣️:

• Common Voice Corpus Delta Segment 18.0 (as of 6/19/2024) 📚.

Background (Noise / Non-speech) 🔇:

• ESC-50 dataset (Harvard Dataverse) 🎧.

(Selected only ~150 folders to keep things manageable.)

• 🧑‍🏫 MLP: trained with Dropout layers to prevent overfitting.
• 🛡️ SVM: trained with LinearSVC from scikit-learn.
• 🔢 LSQ: trained using simple matrix operations.
• 🔁 RNN: trained using SimpleRNN layers to model sequence data.

• 🎯 Tested on three WAV files: 5 seconds, 10 seconds, 20 seconds.
• 📜 Each test file has:
  • .txt file with ground-truth words.
  • .json file with ground-truth timestamps.

Testing Process:

1. Load the test WAV file.
2. Extract Mel spectrogram features.
3. Predict using all models.
4. Post-process with median filtering.
5. Detect speech segments.
6. Compare predictions to ground-truth annotations 📝.

• os 🗂️: File operations
• numpy ➗: Math operations
• json 📄: Handling annotation files
• librosa 🎶: Audio processing
• joblib 💾: Model saving/loading
• scikit-learn 📚: ML algorithms (SVM, preprocessing)
• tensorflow.keras 🤖: Neural networks (MLP, RNN)

• load_train_audio_clips(limit=None): Load training audio.
• extract_features(audio_clip): Get Mel spectrograms.
• pad_features(features, expected_frames): Pad/truncate features.
• Train and save models (MLP, SVM, LSQ, RNN).

• Load audio and ground-truth.
• Predict frame-by-frame speech probability.
• Smooth with median filter.
• Detect segments and compare results.

This project built a speech segmentation system that works without prior word count knowledge.
It uses traditional machine learning and simple RNNs, without heavy neural network models like CNNs, or any external APIs 🌐🚫.