Speech Recognition Project

A student-friendly machine learning project for building real-time voice-to-text and keyword classification workflows — powered by MFCC features and a Random Forest model.

What Is This Project?

This is a student-friendly machine learning project prepared by KKRGENAI for building speech recognition workflows from scratch. No prior ML experience is required — just Python 3.10+ and a working microphone.

You will walk through two complete paths: first, see real-time speech-to-text in action using the Google Web Speech API; then, learn the underlying ML pipeline by recording, preprocessing, and training your own keyword model using MFCC features and a Random Forest classifier.

What You Will Build

Two practical, end-to-end speech recognition paths — both runnable from the command line.

How the Pipeline Works

The keyword recognition path follows a six-stage ML pipeline, each stage handled by a dedicated script in src/.

Project Structure

The repository is organized into clean, purpose-built directories so you always know where each piece lives.

├── data/
│   ├── raw/                  # Student-recorded audio files by label
│   └── processed/            # Cleaned audio and extracted feature files
├── docs/
│   └── student_guide.md      # Beginner step-by-step guide
├── models/                   # Trained model files (.joblib)
├── outputs/                  # Example transcripts and predictions
├── src/
│   ├── collect_audio.py      # Record labeled training audio
│   ├── preprocess_audio.py   # Clean audio & extract MFCC features
│   ├── train_model.py        # Train keyword recognition model
│   ├── predict.py            # Predict keyword from an audio file
│   ├── realtime_transcribe.py# Real-time speech-to-text
│   └── evaluate.py           # Evaluate trained model
├── requirements.txt
└── README.md

Requirements

Step-by-Step Setup

Follow these eight steps to go from a fresh clone to a fully trained keyword recognition model.

1. Create a Virtual Environment

   Isolate your project dependencies from the system Python installation.

   Bash / macOS / Linux

   $ python -m venv .venv
   $ source .venv/bin/activate

   Windows PowerShell

   PS> python -m venv .venv
   PS> .venv\Scripts\Activate.ps1

2. Install Libraries

   Install all required Python packages from the requirements file.

   terminal

   $ pip install -r requirements.txt

   ⚠️ On Windows, if PyAudio fails to install, run:
   pip install pipwin && pipwin install pyaudio then re-run pip install -r requirements.txt.

3. Try Real-Time Voice-to-Text

   Speak clearly after the prompt appears. Your speech will be transcribed and saved automatically.

   terminal

   $ python src/realtime_transcribe.py

   Listening... speak now.
   Recognized text: hello students welcome to speech recognition
   Saved transcript to outputs/live_transcript.txt

4. Record Training Audio

   Create a small labeled dataset. Record at least 5 samples per keyword in a quiet room.

   terminal

   $ python src/collect_audio.py --label start --count 5
   $ python src/collect_audio.py --label stop  --count 5
   $ python src/collect_audio.py --label yes   --count 5
   $ python src/collect_audio.py --label no    --count 5

   Files are saved as: data/raw/start/start_001.wav, data/raw/stop/stop_001.wav, …

5. Preprocess Audio & Extract Features

   Normalize volume, optionally reduce noise, and compute MFCC feature vectors into a CSV.

   terminal

   $ python src/preprocess_audio.py

   Processed 20 audio files
   Saved features to data/processed/features.csv

6. Train the Keyword Recognition Model

   The Random Forest classifier is trained on the extracted MFCC features and saved to disk.

   terminal

   $ python src/train_model.py

   Training samples: 16
   Testing samples: 4
   Accuracy: 0.75
   Saved model to models/keyword_model.joblib

7. Predict a Keyword from Audio

   Point the predictor at any .wav file and get the predicted keyword plus a confidence score.

   terminal

   $ python src/predict.py --audio data/raw/start/start_001.wav

   Predicted keyword: start
   Confidence: 0.86

8. Evaluate the Model

   Generate a full classification report with per-class precision, recall, and F1-score.

   terminal

   $ python src/evaluate.py

   Accuracy: 0.82
   
   Classification report:
   precision recall f1-score support
   no 0.80 1.00 0.89 4
   start 1.00 0.75 0.86 4
   stop 0.75 0.75 0.75 4
   yes 0.80 0.80 0.80 5

Tips for Students

Troubleshooting

$ python -m speech_recognition

$ python -m pip install --upgrade pip setuptools wheel
$ pip install -r requirements.txt

$ python src/collect_audio.py --label start --count 20
$ python src/preprocess_audio.py
$ python src/train_model.py

About KKRGENAI

KKR Gen AI Innovations provides AI-driven solutions, data analytics, software development, app and web development, cloud management, digital services, and AI training for professionals and young learners worldwide.