Kartavya11/language-detection

SmartTTS — Indian Language Identification Model

Model Description

A CNN-LSTM hybrid deep learning model trained to identify spoken Indian languages from audio files. Built as part of the SmartTTS project — an end-to-end multilingual text-to-speech pipeline.

Training Data

• Dataset: Indian Languages Audio Dataset
• Total samples: 10,013 MP3 files
• Duration: 5 seconds per file
• Languages: Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil, Telugu, Urdu

Model Architecture

• 6 × Conv1D layers (pattern detection)
• 3 × MaxPooling layers (dimensionality reduction)
• 2 × LSTM layers (temporal sequence memory)
• 2 × Dense layers (classification)
• Output: Softmax over 11 language classes

Features Extracted (194 total)

• MFCC (40) + Delta MFCC (40)
• Chroma (12) + Mel Spectrogram (40)
• Spectral Contrast (7) + ZCR + Rolloff + RMS

Performance

Metric	Value
Overall Test Accuracy	87.55%
Bengali	99%
English	99%
Gujarati	99%
Malayalam	99%
Marathi	99%
Hindi	97%
Kannada	97%
Tamil	97%
Punjabi	98%
Telugu	93%

How This Model Is Used

This model acts as a quality verification layer in the SmartTTS pipeline:

1. User selects a language and generates speech via Edge TTS
2. This CNN-LSTM model analyses the generated audio
3. Returns confidence score confirming the audio matches the selected language
4. Result is displayed live on the frontend as a verification badge

Live Demo

Try it at: SmartTTS Space

Citation

@misc{smarttts2025,
  author = {Kartavya11},
  title = {SmartTTS Indian Language Identification},
  year = {2025},
  publisher = {HuggingFace}
}