Method of speaker clustering for unknown speakers in conversational audio data

Claim 1

A method for training multiple speaker models based on speech from a multiplicity of speakers, comprising the steps of: a) providing a portion of audio data including speech from at least all of the multiplicity of speakers; b) dividing said portion of audio data into data segments; c) treating said data segments as clusters and computing a pairwise distance between each cluster and every other cluster, basing said pairwise distance on a likelihood measurement that two clusters were created by a same speaker, said likelihood measurement being biased by a Markov duration model based on the prior probability of speaker changes; d) combining two clusters with a minimum pairwise distance into a new cluster.

Claim 2

A method for training multiple speaker models based on speech from a multiplicity of speakers, comprising the steps of: a) providing a portion of audio data including speech from at least all of the multiplicity of speakers; b) dividing said portion of audio data into data segments; c) treating said data segments as clusters and computing a pairwise distance between each cluster and every other cluster, basing said pairwise distance on a likelihood measurement that two clusters were created by a same speaker, said likelihood measurement being biased by a Markov duration model based on the prior probability of speaker changes; d) combining two clusters with a minimum pairwise distance into a new cluster.

Claim 6

A method of indexing audio data according to speaker, said audio data comprising speech from a plurality of speakers, the method comprising the steps of: a) providing a portion of said audio data containing speech from at least all of the plurality of speakers; b) dividing said portion of the audio data into data segments to form clusters; c1) computing a pairwise distance between each cluster and every other cluster, said pairwise distance being based on a likelihood that two clusters were created by a same speaker, said likelihood measurement further being biased by the prior probability of speaker changes; c2) combining a pair of clusters with a minimum pairwise distance; c3) repeating steps c1) and c2) until a desired number of clusters is obtained d) for each of said desired number of clusters, training an individual Hidden Markov Model (HMM) representing a single speaker of the plurality of speakers; e) combining the individual HMMs in parallel to form a speaker network HMM; f) for said audio data, determining a maximum probability path through said speaker network HMM, identifying segments of said audio data associated with each individual HMM; g) marking each segment according to the individual HMM from which it was created; and h) collecting segments according to the marked individual HMM.

Claim 7

A method of indexing audio data according to speaker, said audio data comprising speech from a plurality of speakers, the method comprising the steps of: a) providing a portion of said audio data containing speech from at least all of the plurality of speakers; b) dividing said portion of the audio data into data segments to form clusters; c1) computing a pairwise distance between each cluster and every other cluster, said pairwise distance being based on a likelihood that two clusters were created by a same speaker, said likelihood measurement further being biased by the prior probability of speaker changes; c2) combining a pair of clusters with a minimum pairwise distance; c3) repeating steps c1) and c2) until a desired number of clusters is obtained d) for each of said desired number of clusters, training an individual Hidden Markov Model (HMM) representing a single speaker of the plurality of speakers; e) combining the individual HMMs in parallel to form a speaker network HMM; f) for said audio data, determining a maximum probability path through said speaker network HMM, identifying segments of said audio data associated with each individual HMM; g) marking each segment according to the individual HMM from which it was created; and h) collecting segments according to the marked individual HMM.