ShortTimeObjectiveIntelligibility<T>

Models & Types · AiDotNet.Metrics

Short-Time Objective Intelligibility (STOI) metric for speech intelligibility assessment.

How It Works

STOI predicts the intelligibility of degraded speech signals. It correlates well with human listening tests and is commonly used to evaluate speech enhancement and separation.

Values are correlations, so they lie in [-1, 1]; in practice they fall between 0 and 1, where higher values indicate better intelligibility.

  • 0.9: Excellent intelligibility

  • 0.7-0.9: Good intelligibility
  • 0.5-0.7: Fair intelligibility
  • <0.5: Poor intelligibility

Based on “A short-time objective intelligibility measure for time-frequency weighted noisy speech” by Taal et al. (2011), following the pystoi reference implementation: both signals are resampled to 10 kHz, frames more than 40 dB below the loudest clean frame are removed, the Hann-windowed STFT is grouped into one-third octave bands from 150 Hz, and the clipped, normalized degraded envelopes are correlated with the clean envelopes over 30-frame (384 ms) segments.

When fewer than 30 frames remain after silence removal there is nothing to correlate, and the score is 1e-5 (as in pystoi).

Constructors

Constructor Summary
ShortTimeObjectiveIntelligibility(int sampleRate, int frameLength, int hopLength, int numBands) Initializes a new instance of STOI calculator.

Methods

Method Summary
T Compute(Tensor<T> degraded, Tensor<T> clean) Computes STOI between degraded and clean speech signals.