dwani 0.1.2

Multimodal AI server for Indian languages (speech, vision, LLMs, TTS, ASR, etc.)

Dhwani Server

Dhwani API is a FastAPI-based application providing AI-powered services for Indian languages, including text-to-speech (TTS), language model (LLM) chat, vision-language model (VLM) capabilities, and automatic speech recognition (ASR). It supports lazy loading of models for fast startup and includes endpoints for various tasks.

Features

• Text-to-Speech (TTS): Generate audio from text in Indian languages using Parler TTS.
• Chat: Process Kannada prompts and respond in Kannada via translation and LLM.
• Vision-Language Model (VLM): Caption images, answer visual queries, detect, and point objects.
• Automatic Speech Recognition (ASR): Transcribe audio files in multiple Indian languages.
• Lazy Loading: Models load on-demand or via an explicit endpoint for fast startup.

Prerequisites

• **System Requirements - User **:

  • Python: 3.10
  • Ubuntu 22.04
  • git
  • vscode

• **System Requirements - Server **:

  • Ubuntu with sufficient RAM (16GB+ recommended for models).
  • Optional: NVIDIA GPU with CUDA support for faster inference.

• FFmpeg: Required for audio processing (ASR).

• Server Setup

export HF_HOME=/home/ubuntu/data-dhwani-models
export HF_TOKEN='YOur-HF-token'
python src/server/main.py --host 0.0.0.0 --port 7860 --config config_two

Installation

1. Clone the Repository:

   git clone https://github.com/slabstech/dhwani-server
   cd dhwani-server
   

2. Install Libraries:

   • On Ubuntu: sudo apt-get install ffmpeg build-essential

3. Set Up Virtual Environment:

   python -m venv venv
   source venv/bin/activate 
   

4. Install Dependencies:

   sudo apt-get install -y ffmpeg build-essential
   curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --no-modify-path --profile minimal
   . "$HOME/.cargo/env"
   export CC=/usr/bin/gcc
   export ENV CXX=/usr/bin/g++
   

   pip install --no-cache-dir --upgrade pip setuptools psutil setuptools-rust torch==2.6.0
   pip install --no-cache-dir flash-attn --no-build-isolation
   

   pip install -r requirements.txt
   

5. Set Environment Variable: Create a .env file in the root directory and add your API key: plaintext

   API_KEY=your_secret_key
   

6. Running the Server

   • Start the Server:

       python src/server/main.py --host 0.0.0.0 --port 7860 --config config_two
   

   • The server starts with models loaded on start

   • Access the interactive API docs at http://localhost:7860/docs.

   • (Optional) Load All Models: Preload all models (LLM, Translation, TTS, VLM, ASR) with:

   curl -X POST "http://localhost:7860/load_all_models" -H "X-API-Key: your_secret_key"
   

• Usage

  • Endpoints

    • All endpoints require the X-API-Key header with the value from your .env file.

  • Health Check: GET /health

      curl "http://localhost:7860/health"
  

  • Response:

    {"status": "healthy", "model": "Qwen/Qwen2.5-3B-Instruct"}
    

  • Text-to-Speech: POST /v1/audio/speech

  curl -X POST "http://localhost:7860/v1/audio/speech" -H "X-API-Key: your_secret_key" -H "Content-Type: application/json" -d '{"input": "ನಮಸ್ಕಾರ", "voice": "Female voice", "model": "ai4bharat/indic-parler-tts", "response_format": "mp3"}' --output speech.mp3
  

  • Chat: POST /chat

      curl -X POST "http://localhost:7860/chat" -H "X-API-Key: your_secret_key" -H "Content-Type: application/json" -d '{"prompt": "ನೀವು ಹೇಗಿದ್ದೀರಿ?"}'
  

  • Response: {"response": "<Kannada response>"}
  • Image Captioning: POST /caption/

      curl -X POST "http://localhost:7860/caption/" -H "X-API-Key: your_secret_key" -F "file=@image.jpg" -F "length=short"
  

  • Response: {"caption": "<short caption>"}
  • Visual Query: POST /visual_query/

      curl -X POST "http://localhost:7860/visual_query/" -H "X-API-Key: your_secret_key" -F "file=@image.jpg" -F "query=What is this?"
  

  • Response: {"answer": "<answer>"}
  • Object Detection: POST /detect/

      curl -X POST "http://localhost:7860/detect/" -H "X-API-Key: your_secret_key" -F "file=@image.jpg" -F "object_type=face"
  

  • Response: {"objects": [<list of detected objects>]}
  • Object Pointing: POST /point/

      curl -X POST "http://localhost:7860/point/" -H "X-API-Key: your_secret_key" -F "file=@image.jpg" -F "object_type=person"
  

  • Response: {"points": [<list of points>]}
  • Transcription: POST /transcribe/

  curl -X POST "http://localhost:7860/transcribe/?language=kannada" -H "X-API-Key: your_secret_key" -F "file=@audio.wav"
  

  • Response: {"text": "<transcribed text>"}
  • Batch Transcription: POST /transcribe_batch/

  curl -X POST "http://localhost:7860/transcribe_batch/?language=kannada" -H "X-API-Key: your_secret_key" -F "files=@audio1.wav" -F "files=@audio2.mp3"
  

  • Response: {"transcriptions": ["<text1>", "<text2>"]}

• Notes

  • Lazy Loading: Models load on first use or via /load_all_models. Expect a delay on the first request for each model type. Supported Languages: ASR supports multiple Indian languages (e.g., kannada, hindi, tamil); see models/asr.py for the full list. Logs: Check dhwani_api.log for detailed logs (rotated at 10MB, 5 backups). Performance: Use a GPU with flash-attn installed for faster TTS and ASR inference.

• Troubleshooting

  • Module Errors: Ensure all dependencies are installed. Re-run pip install if needed. FFmpeg Not Found: Install FFmpeg and ensure it’s in your PATH. Permission Denied: Run with sudo if accessing restricted ports (e.g., < 1024).