Systems and methods for fast filtering of audio keyword search

A method for performing an audio keyword search comprising:

• identifying a voice segment of a received audio signal;
• identifying, by a first automatic speech recognition engine, one or more phonemes included in the voice segment;
• outputting, from the first automatic speech recognition engine, the one or more phonemes to a keyword filter;
• receiving a plurality of training speech segments, segment labels associated with the plurality of training speech segments, and a first keyword list including one or more first keywords;
• extracting, by a second automatic speech recognition engine, one or more posteriorgram features associated with the plurality of training speech segments;
• determining, by a machine learning technique, a filter function based on the first keyword list and the one or more posteriorgram features, wherein the filter function enables a binary determination of whether the voice segment includes any of the one or more first keywords of the first keyword list;
• and wherein the filter function collapses frame outputs associated with the voice segment and then estimates phone N-gram counts as a sparse feature vector representation of the voice segment enabling the filter function to make the binary determination;
• executing the filter function to detect whether the voice segment includes any of the one or more first keywords of the first keyword list using the sparse vector representation of the voice segment and, if detected, outputting the one or more phonemes included in the voice segment to a decoder but, if not detected, not outputting the one or more phonemes included in the voice segment to the decoder;
• and
• if the one or more phonemes in the voice segment are output to the decoder:
• generating, by the decoder, a word lattice associated with the voice segment;
• searching the word lattice for one or more second keywords in a second keyword list, and
• determining whether the voice segment includes the one or more second keywords.

A mobile computing device comprising:

• a receiver arranged to receive a communications signal and extract from the communications signal a received audio signal;
• a voice activity detector arranged to identify a voice segment of the received audio signal;
• an automatic speech recognizer including:
• a first automatic speech recognition engine arranged to identify one or more phonemes included in the voice segment and output the one or more phonemes to a keyword filter;
• and
• a decoder arranged to, if the one or more phonemes are outputted by the keyword filter, receive the one or more phonemes included in the voice segment and generate a word lattice associated with the voice segment;
• a keyword search module including:
• a learning module arranged to:
• i) receive a plurality of training speech segments, ii) receive segment labels associated with the plurality of training speech segments, iii) receive a first keyword list including one or more first keywords, and iv) execute a second automatic speech recognition engine to extract one or more posteriorgram features associated with the plurality of training speech segments;
• a keyword classifier arranged to execute a machine learning technique to determine a filter function based on the first keyword list and the one or more posteriorgram features, wherein the filter function enables a binary determination of whether the voice segment includes any of the one or more first keywords of the first keyword list;
• and wherein the filter function collapses frame outputs associated with the voice segment and then estimates phone N-grain counts as a sparse feature vector representation of the voice segment enabling the filter function to make the binary determination;
• and
• the keyword filter arranged to execute the filter function to detect whether the voice segment includes any of the one or more first keywords of the first keyword list using the sparse vector representation of the voice segment and, if detected, output the one or more phonemes included in the voice segment to the decoder but, if not detected, not output the one or more phonemes included in the voice segment to the decoder;
• and
• a word lattice search engine arranged to:
• i) receive the word lattice associated with the voice segment if outputted by the decoder, ii) search the word lattice for one or more second keywords in a second keyword list, and iii) determine whether the voice segment includes the one or more second keywords.