Toolkit developed to aid the process of tracking recording data and ensuring the correct format for training speech enhancement models.
Daily Reports on how I imagined this working and my progress and logic - docuemntation of my process can be fount in the reports folder in this repository.
Recording over 1000 speakers consistently is tougher than it seems. Little differences in volume, file format, or metadata can mess up the training data before the model gets to read it. This toolkit aims to spot those problems early on, at the recording stage, where fixing them is the least troublesome.
- metadata_manager.py — validates and stores recording metadata
- audio_checker.py — batch quality control across a folder of recordings: format consistency, loudness, clipping, silence, and filename validation.
- report_generator.py — put toghetert the metadata and QC results into a summary report (JSON export)
- app.py — desktop UI wrapping all three modules (PyWebView)
the app.py and ui was coded entirely with AI, while on the backend I wrote each definition by myself.
I tested ai|coustics' SDK Playground with a real speech sample. This made me understand why accuracy counts and I instantly thoguh of a tool to aid that process. The model is as good as the recordigns which trained it. The side by side view made that very clear. That's what this tool kit is for. I wanted to understand the problem so I can find a solution at the level of Python knoledge that I have.
pip install -r requirements.txt
python app.pyThe backend logic (metadata_manager.py, audio_checker.py, report_generator.py) was written by me. The UI (app.py) was produced entirely with the help of AI.
Python 3.14 · SQLite · pydub · PyWebView
I am an audio student which has a passion for analysing sound: from speech, to fft algorhythms, to emotion detection, to turning everyday objects and scenes into logical systems. Over the past 7 months I have taugh myself python from scratch (starting with AI assistance, then learning to write it properly myself over the last 4 months) and built tools alongside my other projects. I was an intern and now continue to colaborate now actively with msm-studios and Streamsoft in Berlin and I aspire to learn and actively explore machine learning for audio because that is the direction I plan to take my bachelor thesis.