Systems and methods for voice topic spotting

A voice topic spotting learning system comprising:

• a communications interface arranged to:
• i) receive a plurality of training audio segments and segment topic labels associated with each of the plurality of audio segments, ii) output a fast keyword filter model to a voice topic classifier, and iii) output a topic identification model to the voice topic classifier;
• and
• a processor, in electrical communication with the communications interface, arranged to:
• i) execute an automatic speech recognition engine to extract a plurality of topic keyword hypotheses associated with the plurality of audio speech segments, wherein the automatic speech recognition engine uses a Hidden Markov model to recognize speech within the training audio segments;
• ii) apply a chi-squared test to select a set of topic-indicative words as a subset of the plurality of topic keyword hypotheses and select a training set of topic keyword-containing lattices associated with the set of topic-indicative words;
• iii) generate the fast keyword filter model based on the set of topic-indicative words, and iv) generate the topic identification model based on the training set of topic keyword-containing lattices, where a word posterior indexing function processes the topic keyword-containing lattices to formulate one or more inverted indices to identify the topic associated with the topic keyword-containing audio segment in the topic identification model.

A runtime voice topic spotting classifier system comprising:

• a communications interface arranged to:
• i) receive an audio segment, ii) receive a fast keyword filter model from a voice topic spotting learning system, and iii) receive a topic identification model from the voice topic spotting learning system;
• and
• an automatic speech recognition engine arranged to identify one or more keywords included in a received audio segment and output the one or more keywords, wherein the automatic speech recognition engine uses a Hidden Markov model to recognize speech within the training audio segments;
• a fast keyword filter, implementing the received fast keyword model, arranged to receive the one or more keywords and detect whether the received audio segment includes any keywords of the set of topic-indicative words and, if detected, output the received audio segment as a topic keyword-containing audio segment but, if not detected, not output the received audio segment:
• a decoder arranged to, if the topic keyword-containing audio segment is outputted by the keyword filter, receive the topic keyword-containing audio segment and generate a topic keyword-containing lattice associated with the topic keyword-containing audio segment;
• and
• a voice topic classifier, implementing the received topic identification model, arranged to receive the topic keyword-containing lattice and execute a machine learning technique to determine a topic associated with the topic keyword-containing audio segment, where a word posterior indexing function processes the topic keyword-containing lattice to formulate one or more inverted indices to identify the topic associated with the tonic keyword-containing audio segment in the voice topic identification model.