Simple and basic python function for transcribing audio in realtime based on faster-whisper
This tiny script showcases how easy it is to transcribe audio in realtime (or even faster).
Someday I came across this video: https://youtu.be/Mfbei9I8qQc?feature=shared (btw. a nice channel) when I was looking for a starting point regarding state-of-the-art SpeechToText (STT).
The solution was "faster-whisper": "a reimplementation of OpenAI's Whisper model using CTranslate2, which is a fast inference engine for Transformer models" /1/
For the setup and required hardware / software see /1/
After setting up, feel free to use the code provided in the repo.
Beware: models can use some diskspace, so keep an eye on your system-stats!
important: set your device in line 55/56 (uncomment/comment) and set the correct path in line 56 (ger) or 57 (eng) and choose the target language by commenting/uncommenting one of these lines
After importing necessary libs, of course, we
-
initialize the time-module for time-measurement
-
init the model (Tiny-Whisper, see /2/);
-
feed an audio file (.wav) in "model.transcribe() as well as giving the target language and other inputs (for german you can check e.g. [https://www.lightbulblanguages.co.uk/resources-ge-sound-files.htm])
-
print the transcript and processing-time in terminal-output
important: tested under Win11; Python3.11.5; pip24.0
hardware and setup
CPU: AMD Ryzen 9 7950X GPU: NVIDIA RTX 4090 RAM: 64GB nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2023 NVIDIA Corporation Built on Wed_Nov_22_10:30:42_Pacific_Standard_Time_2023 Cuda compilation tools, release 12.3, V12.3.107 Build cuda_12.3.r12.3/compiler.33567101_0
-
read https://github.com/SYSTRAN/faster-whisper?tab=readme-ov-file#requirements /1/
-
create a venv (Win11, e.g.:
python -m venv stt-venv) -
activate venv (Win11:
stt-venv/Scripts/activate) -
pip install -r requirements.txt -
run
python stt.py -
have fun and feel free to share your use-cases :)
- feel free to add functionality
- I use it for a STT-LLM-TTS pipeline based on "faster-whisper" (STT), Ollama & API (LLM, Mistral 7b), suno-bark (TTS)