Recognition-processing systems for speech enhancement: application to frequency lowering

A method is presented for recoding high-frequency speech components into a low-frequency region, to increase audibility for certain sensorineural-loss listeners. The method, called recognition processing (RP), is based on a combination of disparate earlier work on frequency lowering using real-time speech processing and low-frequency signal synthesis. The RP system produced 800-Hz lowpass-filtered speech augmented with additive low-frequency noise signals during obstruent segments. Different low-frequency noise properties were chosen for each obstruent type; the noise signals were introduced at times determined by prior time-aligned phonemic transcription. Normal-hearing subjects listened to CV syllables processed by this technique and using a reference system. RP generally resulted in slightly higher identification scores