Uktam.ai

Uktam.ai is a powerful, entirely offline Android application for real-time speech-to-text (ASR) transcription and machine translation between Indic languages. Built with modern Android development practices, it runs state-of-the-art AI models directly on your device—ensuring complete privacy and zero reliance on cloud APIs.

      
Real-time, fully offline Indic speech-to-speech linguistic pipeline running directly on your Android device.

Why Uktam.ai?

• 100% Offline & Private: No audio recordings or text transcripts ever leave your device. Every single piece of processing—from speech recognition to large language model translation—happens locally on your smartphone's silicon.
• Made for India, by Indian AI: Instead of relying on generic global models, Uktam.ai is strictly built around models researched and trained specifically for Indian languages. By utilizing Sarvam AI and AI4Bharat, the app captures the nuances, dialects, and grammatical complexities of languages far better than generic cloud APIs.
• Custom Quantized for Mobile: Running multi-billion parameter AI models on a phone requires immense compute. Uktam.ai uses models that were custom-quantized (GGUF/ONNX) specifically for this project. This hands-on optimization drastically reduces the memory footprint and battery consumption, allowing them to run smoothly on standard smartphone hardware while maintaining near-perfect accuracy.
• Zero Latency / No Internet Required: Perfect for remote areas, low-connectivity zones, or traveling. Once the models are downloaded, you never need Wi-Fi or mobile data to translate again.

Key Features

• Supported Languages: Currently supports real-time speech recognition, translation, and text-to-speech between Hindi, Kannada, Tamil, Telugu, Marathi and Malayalam (more Indic languages coming soon!)
• Real-time Offline Speech Recognition: Powered by Sherpa-ONNX using the AI4Bharat IndicConformer model for high-speed, local transcription.
• On-Device Translation: Leverages the Sarvam Translate model (from Sarvam AI) running via llama.cpp through JNI for highly accurate offline machine translation.
• Native Text-to-Speech (TTS): Automatically speaks the translated text using Android's native offline TTS engine.
• Premium UI/UX: Built entirely with Jetpack Compose, featuring:
  • Sleek, high-contrast Dark and Light modes.
  • Haptic feedback for tactile interactions (Vibration and TouchApp integrations).
  • Smooth transition animations and dynamic chat bubbles.
• Intelligent Model Management: ASR models are bundled seamlessly via Play Asset Delivery. For the translation model, a built-in downloader intelligently fetches the optimal llama.cpp model size based on your device's available RAM to prevent memory crashes.

Technology Stack

• Language: Kotlin, C++
• UI Framework: Jetpack Compose, Material Design 3
• Architecture: MVVM (Model-View-ViewModel) with StateFlow and Coroutines.
• AI & Machine Learning:
  • Sherpa-ONNX for Automatic Speech Recognition (ASR) running the AI4Bharat IndicConformer.
  • llama.cpp (via custom JNI bindings) running Sarvam Translate for offline translation.
• Build System: Gradle (Kotlin DSL), CMake for native C++ compilation.

Getting Started

Prerequisites

• Android Studio (Koala or newer recommended)
• Android SDK Minimum API Level: 34 (Android 14)
• Minimum Device RAM: 6GB Recommended. The app features dynamic model selection: devices with >6GB RAM receive a higher-accuracy model (Q4_K_S), while devices with 6GB or less download a smaller, more efficient model (Q2_K) to prevent memory crashes.
• Android NDK & CMake (Required for building the llama.cpp JNI bindings)

Installation & Setup

1. Clone the Repository

   git clone https://github.com/ashb155/uktam.git
   cd uktam

2. Sync Gradle & NDK Open the project in Android Studio. Ensure that your SDK Manager has the NDK (Side by side) and CMake installed. Gradle will automatically sync and build the C++ bindings via the CMakeLists.txt.

3. Model Preparation The application requires the ONNX models for ASR and the .gguf file for Llama.

   • The ASR models are bundled via Play Asset Delivery (:asr_assets) and are installed automatically.
   • Upon launching the app for the first time, the DownloadScreen will fetch the Sarvam Translate model (approx. 1GB to 2.5GB).
   • Ensure you have an active internet connection for this initial step so the app can download the required .gguf model to your device's internal storage.

4. Run the App Connect a physical Android device (emulators may struggle with local model inference without hardware acceleration) and click Run.

Contributing

Contributions, issues, and feature requests are welcome! If you plan to implement major features (such as adding support for new Indic languages), please open an issue first to discuss the proposed changes.

License

This project is licensed under the GPL-3.0 License.