A binaural short time objective intelligibility measure for noisy and enhanced speech

Asger Heidemann Andersen, Jan Mark de Haan, Zheng-Hua Tan, Jesper JensenInterspeech 2015 · DOI

Introduction

Understanding speech in noisy environments is a complex task that humans perform remarkably well thanks to binaural hearing—the ability to process sounds from both ears. For engineers designing hearing aids and speech-enhancement algorithms, predicting how well a user will understand processed audio is crucial. While the Short-Time Objective Intelligibility (STOI) measure has become a standard for monaural (single-ear) audio, it fails to account for the "binaural advantage" where spatial separation between speech and noise improves clarity.

Researchers have introduced BSTOI (Binaural Short-Time Objective Intelligibility), a new metric designed to bridge this gap. Unlike previous binaural models that struggle with non-linear signal processing, BSTOI accurately evaluates "enhanced" speech, making it a vital tool for developing next-generation communication devices.

The BSTOI Model

The core innovation of BSTOI lies in its ability to account for both binaural hearing and the distortions introduced by modern noise-reduction algorithms. The model combines the STOI framework with a modified Equalization-Cancellation (EC) stage.

Block diagram of the BSTOI model showing the three-step process: time-frequency decomposition, modified EC-stage, and intelligibility prediction.

The process follows three main steps:

  1. Time-Frequency Decomposition: Input signals (clean reference and noisy/processed speech for both ears) are decomposed into time-frequency tiles using a Short-time Discrete Fourier Transform (DFT).
  2. Modified EC Processing: This stage simulates how the brain aligns signals from both ears to minimize noise. Traditional EC models maximize the Signal-to-Noise Ratio (SNR). However, since noise in processed speech is often inseparable from the signal, BSTOI finds parameters (time-shift and level-adjustment) that maximize the STOI correlation itself. Internal noise is added to these parameters to mimic biological limitations.
  3. Intelligibility Prediction: The optimized binaural stream is compared against the clean reference across 15 third-octave bands. The resulting correlation coefficient (between 0 and 1) represents the predicted intelligibility.

Validation and Results

The model was validated using diverse datasets, including the Dantale II sentence corpus and the CIPIC database for realistic spatial audio simulation. A key part of the evaluation involved testing BSTOI against human performance in scenarios with complex noise and beamforming processing.

Comparison of measured vs. predicted intelligibility across various noise conditions and processing techniques, showing a high correlation (ρ = 0.974).

Key findings include:

  • High Accuracy: BSTOI showed a very high correlation with human performance ($\rho = 0.974$) across conditions involving isotropic noise, point-source interferers, and speech-like noise.
  • Binaural Advantage: The model accurately predicted "spatial release from masking." When noise sources were moved away from the target speaker, the predicted Speech Reception Thresholds (SRTs) matched human data with a maximum error of less than 0.7 dB.
  • Robustness: BSTOI successfully handled non-linear processing like Ideal Time-Frequency Segregation (ITFS) and MVDR beamforming, maintaining its predictive power where older metrics failed.

Limitations and Future Outlook

While BSTOI represents a significant advancement, it currently focuses on anechoic environments. Future research will extend the model to reverberant rooms, which present a more challenging binaural listening scenario. Additionally, while the current model is tuned for normal-hearing listeners, adjusting the "internal noise" parameters in the EC stage could allow BSTOI to predict intelligibility for hearing-impaired users, further aiding the development of personalized hearing technology.

Summary updated

BibTeX
@inproceedings{andersen2015binaural,
  author = {Asger Heidemann Andersen and Jan Mark de Haan and Zheng-Hua Tan and Jesper Jensen},
  title = {{A binaural short time objective intelligibility measure for noisy and enhanced speech}},
  booktitle = {Interspeech},
  year = {2015},
  doi = {10.21437/interspeech.2015-552},
}
Figure