SYSTRAN/faster-whisper ? reverse-engineered prompt

I need a Python library for speech-to-text transcription. It should be based on OpenAI's Whisper model, but an optimized version that runs much faster—the one I saw was up to 4 times faster—and uses way less memory.

The main aility should be to take an audio file like an MP3, transcribe it, and then output the text broken down into segments with start and end timestamps. It also needs to be able to detect what language is being spoken and tell me what it found.

Please set it up so it can run on a normal CPU, but also make it easy to switch to a GPU for even more speed if I have an NVIDIA card. There should also be an option to use a lower-precision mode (like int8 quantization) to make it even more efficient. The installation should be simple via pip, and it shouldn't require me to install FFmpeg separately. Finally, include an example of how to load a model and transcribe a file.