Acoustic model training method and apparatus

摘要: The invention discloses an acoustic model training method and apparatus. One specific embodiment of the includes steps: removing high delay search paths among all during process by means a connectionist temporal classification criterion, wherein are that output state is greater than threshold; based on less threshold, model. apparatus realize CTC (connectionist classification) thesearch removed so as to enable avoid taking part in model, thus avoiding problem sequence from after has hysteresis quality because large quantity take when criterion used train enabling have lower time predicting voice.