Objective Speech Intelligibility Measure for Low Bit-Rate Speech Codecs Operating in Noisy Channels

Low bit-rate speech codecs are important in many critical communication systems used by the military, police and fire departments, homeland security and other first responders. Such codecs offer low bandwidth, ease of encryption, and reasonable voice quality. Their increasing role in emergency communication scenarios requires them to function reliably under marginal RF channel conditions exhibiting low signal-to-noise ratios and multi-path in a harsh mobile environment. The result is increased bit error rates that may seriously degrade speech intelligibility. This paper introduces an objective tool based on a statistical language model for assessing intelligibility in these situations. Such a tool could be used to develop communication systems that are robust in harsh RF channels, for evaluating the effectiveness of radios prior to deployment, and perhaps even for real-time monitoring of system performance. Subjective human listener-based assessment is always the gold standard; however, the time and expense of such tests make them inappropriate for many practical applications. For the proposed technique, correlation with subjective intelligibility was measured to be 0.96, and appears to be text, speaker, and technology independent.