Reference-free measurement of dialog intelligibility in (post)-production, content analyses, and live broadcast applications

Abstract

Speech intelligibility and listening effort are receiving growing interest in broadcast. Since extensive expert assessments and formal listening tests are too costly in practice, instrumental measures to analyze intelligibility are required. Existing measures include energy-based methods (e.g., speechto-background loudness differences), and auditory model-based methods, which mimic the auditory system and derive their assessment based on a comparison of how well important speech features are preserved in a mix. A third class of measures makes us of advances in automatic speech recognition technologies. These approaches have recently been shown to generalize well to speech-impairing factors which had not been included during training, suggesting that they might generalize well and be applicable to a variety of listening conditions. This may be an advantage in broadcast, where there are many different potential sources of poor intelligibility which are difficult to capture with traditional metrics. This contribution discusses potentials, limitations, and possible applications of current intelligibility measures.