Transcribe desktop audio/computer audio in real-time and locally (Streaming ASR), using TorchAudio and Emformer-RNNT model.
- Reimplemented stream reader using PyAudio rather than
torchaudio.io. Thus no need for ffmpeg. - Asynchronous data transfer via a thread-safe queue.
- Minimalist GUI to display transcriptions, built with Tkinter.
There are great models like OpenAI's Whisper to transcribe audio files, but they can't do it in real-time.
There are some apps which transcribe microphone audio, but they can't transcribe desktop audio.
This project, however, transcribes any video or audio which is playing on your computer in real-time.
But this project isn't perfect. There is only one available model: EMFORMER_RNNT_BASE_LIBRISPEECH, which only supports English. And it's accuracy isn't satisfying.
But still, it demonstrated one way to do such things.
