This project provides a self-hostable API for speech-to-text transcription using a finetuned Whisper ASR model. It targets developers needing to integrate speech recognition into their applications, offering user-level access via API keys and optimized inference for efficient processing.

How It Works

The API leverages a finetuned and quantized Whisper ASR model for accurate and fast speech-to-text conversion. It exposes a simple HTTP interface, allowing users to upload audio files and receive transcriptions. The use of quantized models aims to reduce resource consumption and improve inference speed, making it suitable for self-hosting and integration.

Quick Start & Requirements

• Install ffmpeg: sudo apt install ffmpeg
• Install Python dependencies: pip install -r requirements.txt
• Run the project: uvicorn app.main:app --reload
• Get API token: POST request to /api/v1/users/get_token with email and password.
• Transcribe audio: POST request to /api/v1/transcribe/ with model query parameter and audio file upload.
• Available models include tiny.en and quantized versions like tiny.en.q5.
• Documentation: https://github.com/innovatorved/whisper.api

Highlighted Details

• Finetuned Whisper ASR model for enhanced accuracy.
• Quantized model optimization for faster, efficient inference.
• User-level access control with API keys.
• Self-hostable architecture for full control.

Maintenance & Community

• Primary author: Ved Gupta.
• Support contact: vedgupta@protonmail.com.

Licensing & Compatibility

• License: MIT.
• Permissive license allows for commercial use and integration into closed-source applications.

Limitations & Caveats

The provided example for obtaining a token uses placeholder credentials, and users are instructed to use tokens provided by an admin, implying a managed deployment or setup process for token generation.