Download Whisper Tarteel model:
https://huggingface.co/tarteel-ai/whisper-base-ar-quran
Install Python Requirements
pip install -r requirements.txt
Run Flask Server:
python3 local_app.py
Alternatively, if you set up a OpenAI key, you can do inference with the hosted version of Whisper by setting your environment variable using this guide and then running python3 hosted_app.py
Navigate to http://127.0.0.1:5000 to use the app!
- Split up audio into chunks for processing in parallel (every 3 seconds)
- Add a event / session / recording model to keep track of audio clips and join if necessary
- On html page, add separate tracks per ayat maybe
- Remove background noise and parse out non-arabic sections
- Merge outputs together into a summary of what was recited
- Add UI/UX for easy use
- Set up server
- Change to a word based algorithm, from a passage based algorithm to account for repetitions
- Quantize model https://pytorch.org/tutorials/intermediate/dynamic_quantization_bert_tutorial.html#evaluate-the-inference-accuracy-and-time https://huggingface.co/docs/optimum/concept_guides/quantization
- Try out Faster-Whisper (https://github.com/guillaumekln/faster-whisper)
- Word based approach could start using string matching for different roots for bi-grams, tri-grams etc
- Deal with a tree of expanding notes of roots
- stop recording after certain amount of silence 1 minute or so
- Fix the issue when users repeat a verse
- Potential solution: Remove repetitious phrases
- Look into switching to streamlit
- Add tests and componentize functions
- Whisper model to transcribe speech to arabic using Tarteel AI
- Once transcript is retrieved, we use rapidfuzz for the string matching