Rejection method for low-resource embedded speech recognition

The invention relates to a rejection method for low-resource embedded speech recognition. The rejection method includes the steps: constructing a command word decoding network and a filler decoding network based on a command word list and a filler list respectively, and parallelly connecting the filler decoding network into the command word decoding network so as to form a complete decoding network space; generating a speech background probability score list; inputting speech to be identified into the decoding network space for primary recognition; performing decoding recognition: rejecting filler in a primary recognition result; and performing secondary discrimination on command words in the primary recognition result: performing confidence degree discrimination so as to obtain a final recognition result. According to the rejection method for low-resource embedded speech recognition, out-of-vocabulary word rejection rate can be increased, memory usage is reduced and application requirements of low-resource condition embedded speech recognition are met.