demo/

Inference-side demo for the Spoken Term Detection (STD) system.

Files

File	Purpose
build_dbase_index.py	Build the retrieval DB, TF-IDF matrix, and FAISS index for one (encoder, codebook_size, split). Run once before searching.
search_clean.py	Evaluate STD on clean queries. IV/OOV chosen via MANUAL TOGGLE inside the file.
search_noise.py	Evaluate STD on noise-corrupted queries, sweeping SNR ∈ {−5, 0, 5, 10, 15, 20} dB. Same MANUAL TOGGLE as above.
extract_token_sequences_for_word_pairs.ipynb	Notebook: tokenise same-word utterance pairs and compute jaccard similarity.

Usage

# 1. Build the retrieval index 
python build_dbase_index.py --split test-clean --codebooksz 4096

# 2a. Search with clean queries
python search_clean.py --split test-clean --codebooksz 4096 --encoder bimamba

# 2b. Search with noisy queries (SNR sweep)
python search_noise.py --split test-clean --codebooksz 4096 --encoder bimamba

To switch between IV and OOV queries, grep for MANUAL TOGGLE inside the search scripts and edit both marked lines.

Asset paths (/home/anup/..., /DATA/...) and CUDA_VISIBLE_DEVICES are hardcoded at the top of each script — edit before running locally.