Voice segment detection for extraction of sound source

A voice recognition method, comprising:

• detecting, by a tracking unit, a voice segment from a plurality of voice segments of an observation signal input from a microphone;
• setting, by the tracking unit, a first registered attribute to identify the voice segment based on an explicit sign of a speaker, wherein the explicit sign is provided based on an image analysis;
• creating, by the tracking unit, a management unit;
• setting, by the management unit, the voice segment by connection of a detection result of the voice segment to a time direction;
• detecting, by the management unit, a sound source direction of a sound source;
• updating, by the management unit, the set voice segment based on the detection of the sound source direction;
• creating, by the management unit, an extraction filter;
• extracting, by the extraction filter, the sound source after a time has elapsed from a beginning of the voice segment;
• applying, by the management unit, the extraction filter to the observation signal;
• changing, by the management unit, a mask that decreases a transmittance of the observation signal for each frequency and each time, wherein the mask is changed when a proportion of a sound other than a target sound is higher than the target sound in the observation signal;
• executing, by the management unit, a time frequency masking process for application of the changed mask to the observation signal;
• extracting, by the management unit, the sound source based on the execution of the time frequency masking process;
• creating, by the management unit, a partial sound source extraction result of the voice segment, based on the application of the extraction filter to the observation signal and the extraction of the sound source by the execution of the time frequency masking process;
• outputting, by the tracking unit, the partial sound source extraction result to a voice recognition unit;
• inputting, by the voice recognition unit, the partial sound source extraction result from the tracking unit;
• executing, by the voice recognition unit, a voice recognition process on the partial sound source extraction result;
• outputting, by the voice recognition unit, a voice recognition result based on the voice recognition process;
• determining, by a ranking unit, a priority of the voice segment of the plurality of voice segments;
• and
• outputting, by the ranking unit, a plurality of sound source extraction results to the voice recognition unit based on the determined priority.

A non-transitory computer-readable medium having stored thereon computer-readable instructions, which when executed by a computer, cause the computer to execute operations, the operations comprising:

• detecting a voice segment from a plurality of voice segments of an observation signal input from a microphone;
• setting a first registered attribute to identify the voice segment based on an explicit sign of a speaker, wherein the explicit sign is provided based on an image analysis;
• creating a management unit;
• setting the voice segment by connection of a detection result of the voice segment to a time direction;
• detecting a sound source direction of a sound source;
• updating the set voice segment based on the detection of the sound source direction;
• creating an extraction filter;
• extracting the sound source after a time is elapsed from a beginning of the voice segment;
• applying the extraction filter to the observation signal;
• changing a mask that decreases a transmittance of the observation signal for each frequency and each time, wherein the mask is changed when a proportion of a sound other than a target sound is higher than the target sound in the observation signal;
• executing a time frequency masking process for application of the changed mask to the observation signal;
• extracting the sound source based on the execution of the time frequency masking process;
• creating a partial sound source extraction result of the voice segment, based on the application of the extraction filter to the observation signal and the extraction of the sound source by the execution of the time frequency masking process;
• outputting the partial sound source extraction result to a voice recognition unit;
• inputting the partial sound source extraction result;
• executing a voice recognition process on the partial sound source extraction result;
• outputting a voice recognition result based on the voice recognition process;
• determining a priority of the voice segment of the plurality of voice segments;
• and
• outputting a plurality of sound source extraction results to the voice recognition unit based on the determined priority.