[07 // Machine Learning]
Speech-to-Text System
- Python
- Wav2Vec2
- Hugging Face Transformers
- Librosa
- FFmpeg
- Pandas
[01]
Overview
This Google Colab workflow converts audio or video into reusable transcript and subtitle outputs.
It normalizes source media, performs inference with the pretrained facebook/wav2vec2-base-960h model, and exports TXT, CSV, JSON, SRT, and burned-in subtitles.
[02]
My Contribution
- Built the media-ingestion flow for audio and video sources.
- Implemented mono 16 kHz conversion and chunk-based processing before inference.
- Structured transcript output for plain text, tabular, JSON, and subtitle formats.
- Added SRT generation and an FFmpeg step for burned-in subtitle output.
[03]
Key Technical Notes
- Audio is converted to mono at a 16 kHz sample rate before model inference.
- Long inputs are divided into chunks so the workflow can process them sequentially.
- The implementation uses the pretrained Wav2Vec2 base 960-hour model as documented.
- The available project record does not include a benchmark evaluation.
[04]
Visual Evidence
[05]
Tech Stack
- Python
- Wav2Vec2
- Hugging Face Transformers
- Librosa
- FFmpeg
- Pandas
[06]
Status & Link
Completed workflow
Source repository or production deployment is private.