isaacnetero/polyvox-kpi-evaluation

PolyVox KPI Evaluation Dataset

This dataset supports PolyVox evaluation for Problem Statement 11: Real-Time Multi-User Smart Assistant for Dynamic and Noisy Smart Environments.

It contains compact evaluation mixtures for 2-speaker and 3-speaker speech separation under clean, noisy, overlap, and optional RIR-style conditions.

Contents

• manifests/kpi_dataset_manifest.csv
• dataset_summary.json
• audio/ containing mixtures and clean source references

Source Data And Attribution

This dataset is derived from or inspired by:

• LibriSpeech, CC BY 4.0: https://www.openslr.org/12/
• MUSAN, CC BY 4.0: https://www.openslr.org/17/
• LibriMix metadata/recipe: https://github.com/JorisCos/LibriMix
• SNIPS/Sonos NLU benchmark, CC0-1.0: https://github.com/sonos/nlu-benchmark

Intended Use

This dataset is intended for reproducing PolyVox separation and realtime-assistant evaluation results. It is not a replacement for the original source corpora.

Problem Statement KPIs

The dataset is designed to evaluate:

• 2-speaker and 3-speaker separation
• clean and noisy mixtures
• SI-SNR
• WER
• real-time factor
• model behavior under overlapping speech