For the DARPG Hackathon, The Problem Statement 3 involved evaluating and optimizing an Open Source speech-to-text model to accurately transcribe feedback calls related to citizen grievances into English text.
Since the textual output data was not provided, Whisper LLM was used to generate textual data for each audio dataset. This data was then stored in a metadata.csv file, and after preprocessing, it was used to fine-tune the Whisper small LLM.
Prepare your Audio folder in the following format:
audio_dataset/
├── metadata.csv
└── data/
metadata.csv contains the names of the audio files audio_path and their corresponding texts transcription.
I have pushed the fine-tuned model to 🤗Hugging Face under the name sanket003/whisper-darpg.
To use the model, run the run_model.py script, which contains a Gradio interface for easy interaction with the model.