Scale-invariant signal-to-distortion ratio improvement (SI-SDRi)

Scale-invariant signal-to-distortion ratio improvement (SI-SDRi) is a metric used to evaluate the performance of speech and audio source separation algorithms. It measures the improvement in the quality of the separated signal over the original mixed signal, taking into account the scale-invariance of audio signals.

The SI-SDRi is defined as the ratio between the scale-invariant signal-to-distortion ratio (SI-SDR) of the separated signal and the SI-SDR of the mixed signal, both measured in decibels (dB). The SI-SDR is a variant of the signal-to-distortion ratio (SDR) that takes into account the scaling and permutation of the separated sources. It is computed as the ratio between the energy of the reference source and the energy of the distortion signal, after aligning the sources using the permutation invariant training (PIT) method.

The SI-SDRi is a more robust and meaningful metric than the traditional SDR, especially for scenarios where the scaling and permutation of the sources are unknown or arbitrary. A higher SI-SDRi value indicates a better separation performance, with a value of 0 dB indicating no improvement and higher values indicating greater improvement.

Trustworthy AI Relevance

This metric addresses Robustness and Transparency by quantifying relevant system properties. Robustness: SI-SDRi quantifies how much a model improves signal fidelity under noisy or interfering conditions (scale-invariant), making it directly useful for assessing noise resilience, degradation under distribution shift, and comparative reliability across models and conditions. Tracking SI-SDRi across environments and inputs helps identify failures and measure robustness-improving interventions. Transparency: As a standardized, interpretable scalar metric, SI-SDRi supports transparent reporting and comparison of audio-model performance.