🎙️ Audio Language Detection

AI-powered spoken-language classification for Indian languages using acoustic features and Random Forest.

Live Demo: https://audio-language-detection-bhwha53elusztx56o8yb2v.streamlit.app/

---

✨ Overview

Audio Language Detection is a Streamlit application that analyzes spoken audio and predicts which supported Indian language is being spoken.

The pipeline combines signal preprocessing, handcrafted acoustic feature extraction, feature scaling, and a trained Random Forest classifier. The application supports both uploaded audio files and direct microphone recording, then presents the prediction together with confidence-oriented visualizations.

Supported Languages

Language	Supported
🇮🇳 Bengali	✅
🇮🇳 Gujarati	✅
🇮🇳 Kannada	✅
🇮🇳 Konkani	✅
🇮🇳 Malayalam	✅
🇮🇳 Marathi	✅
🇮🇳 Odia	✅
🇮🇳 Tamil	✅
🇮🇳 Telugu	✅

---

🚀 Live Application

Open the deployed application:

https://audio-language-detection-bhwha53elusztx56o8yb2v.streamlit.app/

You can upload a supported audio file or record speech directly from the browser and inspect the model output and visualizations.

---

🎯 Key Features

🎧 Flexible Audio Input

• Upload .wav, .mp3, .m4a, .ogg, or .flac audio files.
• Record speech directly from the browser microphone.

🧹 Audio Preprocessing

The inference pipeline prepares incoming audio through:

• Resampling to 16 kHz
• Spectral noise reduction with noisereduce
• DC-offset correction
• Silence trimming / VAD-style cleanup

🧠 78-D Acoustic Representation

The classifier uses a 78-dimensional acoustic feature representation built from:

• MFCCs
• MFCC deltas
• Chroma features
• Spectral contrast
• Tonnetz
• Spectral centroid
• Zero-crossing rate

📊 Model Output & Visualization

• Predicted language
• Probability distribution across supported classes
• Audio waveform
• Mel-spectrogram
• Interactive exploratory visuals

---

🏗️ Pipeline

Audio File / Microphone
          │
          ▼
   Audio Loading
          │
          ▼
 Resampling + Cleanup
          │
          ▼
Noise Reduction + Trimming
          │
          ▼
  78-D Feature Extraction
          │
          ▼
       Scaling
          │
          ▼
 Random Forest Classifier
          │
          ▼
Language Prediction
          │
          ├──────────────► Confidence Distribution
          ├──────────────► Waveform
          └──────────────► Mel-Spectrogram

---

📁 Project Structure

.
├── page.py
├── inference.py
├── train_and_save_model.py
├── model_bundle.joblib
├── requirements.txt
├── sample_audios/
├── colab_notebook.ipynb
└── work/

Core Files

File	Purpose
page.py	Streamlit interface for upload, recording, prediction, and visualization
inference.py	Audio preprocessing and 78-D feature extraction used during inference
train_and_save_model.py	Model training and bundle generation
model_bundle.joblib	Serialized model artifacts used for inference
requirements.txt	Python dependency list
sample_audios/	Sample/benchmark audio clips
colab_notebook.ipynb	Experimental and research workflow
work/	Benchmark/evaluation artifacts

---

⚡ Quick Start

1. Clone the repository

git clone https://github.com/RICK2814/audio-language-detection.git
cd audio-language-detection

2. Install dependencies

pip install -r requirements.txt

3. Run the Streamlit application

python -m streamlit run page.py

Then open:

http://localhost:8501

---

🧪 Inference Flow

The application follows the same general path for uploaded audio and microphone recordings:

1. Load the audio signal.
2. Normalize the sampling rate to 16 kHz.
3. Reduce unwanted spectral noise.
4. Correct DC offset and trim silence.
5. Compute the 78-D acoustic representation.
6. Apply the trained preprocessing/scaling pipeline.
7. Run the Random Forest classifier.
8. Decode the predicted class label.
9. Visualize prediction confidence and audio characteristics.

---

🤖 Machine Learning Architecture

The repository stores a pre-trained inference bundle containing the artifacts needed to transform extracted features into the final language prediction.

Conceptually:

Raw Audio
   ↓
Acoustic Feature Engineering
   ↓
Feature Scaling
   ↓
Random Forest
   ↓
Label Decoding
   ↓
Predicted Indian Language

This design keeps feature extraction and model inference separated from the Streamlit presentation layer, making the system easier to test and extend.

---

📈 Visual Analysis

The application is designed not only to return a class label but also to expose interpretable audio-side information through visualizations such as:

• Waveform — amplitude variation over time
• Mel-spectrogram — time-frequency representation on a mel scale
• Class probability chart — relative model confidence across supported languages

---

🔬 Research / Experimentation

The repository includes a Colab notebook and working artifacts for experimentation and benchmark analysis. These materials can be used to inspect the feature pipeline, training workflow, and evaluation process before deploying changes to the Streamlit interface.

---

🌐 Deployment

The application is deployed on Streamlit Community Cloud.

For a local deployment:

python -m streamlit run page.py

For Streamlit Community Cloud, connect the GitHub repository and configure the application entry point as:

page.py

Install dependencies from:

requirements.txt

---

🛠️ Development Commands

# Install dependencies
pip install -r requirements.txt

# Run locally
python -m streamlit run page.py

# Train / rebuild the model bundle
python train_and_save_model.py

The training command should only be run when you intentionally want to regenerate the model artifacts used by the application.

---

✅ Recommended Test Checklist

Before publishing a new version, verify:

• App launches locally
• .wav upload works
• .mp3 upload works
• Microphone recording works
• Preprocessing completes without errors
• All supported language labels decode correctly
• Probability visualization renders
• Waveform renders
• Mel-spectrogram renders
• model_bundle.joblib is available in the deployment environment
• Streamlit deployment starts successfully

---

⚠️ Notes & Limitations

Prediction quality depends on recording conditions, speaker characteristics, pronunciation, background noise, sample duration, and how representative the training data is of real-world speech.

A model confidence score should be interpreted as the classifier's estimated class probability distribution, not as a guarantee of correctness.

---

🔐 Model & Data Handling

The application performs local preprocessing of the supplied audio before inference. The repository includes the serialized model bundle used for prediction. Avoid committing private or sensitive recordings to the repository.

---

📦 GitHub Repository

Source: https://github.com/RICK2814/audio-language-detection

Live Demo: https://audio-language-detection-bhwha53elusztx56o8yb2v.streamlit.app/

---

🗺️ Roadmap

Potential next improvements:

• Add more Indian languages
• Add stronger calibration and evaluation reporting
• Improve robustness to noisy and short recordings
• Add richer model benchmarking
• Add confusion-matrix and per-class evaluation views
• Introduce automated testing for preprocessing and inference
• Add model version metadata to the UI

---

📄 License

This project is released under the MIT License.

---

⭐ Support the Project

If this project is useful for speech-processing, machine-learning, or Indian-language research, consider starring the repository and sharing feedback.