Speaker Segmentation and Clustering using Gender Information

Abstract

Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searciing existing data sources, gathering and maintaining the data needed, and completing and reviewng this collection of information. Send comments regarding this burden estimate or any other aspect of this collection of irformation, including sugglestions for reduang this burden to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson Daws H-ihway, Suite 1204, Arlington. VA 22202-4302. Rfespondont should be &Aare that notwithstanding any other provision of law, no person a"e be subject to any penalty for failing to comply with a collection of information if a does not display a Currently vald OMB control number PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS. 14. ABSTRACT This paper considers the segmentation and clustering of conversational speech for the two-wire training (3conv2w) and two-wire testing (lconv2w) conditions of the NIST 2005 Speaker Recognition Evaluation. A notable feature of the system described is that each file is labeled as containing either opposite-or same-gender speakers. The speech segments for opposite-gender files are clustered by gender, while those for same-gender files are processed by agglomerative clustering. By using gender information in the clustering of the opposite-gender files, the equal error rate in the 3conv2w training condition was reduced from 15.2% to 9.9%. For the lconv2w testing condition, clustering opposite-gender files by gender did not improve performance over agglomerative clustering; however, it was over 100 times faster than agglomerative clustering on the opposite-gender files. [4], is to use the Bayesian InformaThis paper considers the segmentation and clustertion Criterion (BIC) to detect change points. This method ing of conversational speech for the two-wire training is based on modeling the audio stream as a Gaussian pro-(3eonv2w) and two-wire testing (lconv2w) conditions cess and using a maximum likelihood approach to detect of the NIST 2005 Speaker Recognition Evaluation. A potential changes. There have been numerous adaptanotable feature of the system described is that each file tions of the BIC procedure, including the DISTBIC proisolabeled far con g ethe r oppoemdescrite-or shameendhler cedure of [2] and the modified BIC procedure of [5] that is labeled as containing either opposite-or same-gender removes the necessity for having to choose the threshold are clustered by gender, while those for same-gender files in the penalty term. In [6,7], gender and channel detecare processed by agglomerative clustering. By using gention have been used in the first stages of segmentation for der information in the clustering of the opposite-gender speaker diarization of news broadcasts. files, the equal error rate in the 3conv2w training condiClustering can be defined as grouping homogeneous tion was reduced from 15.2% to 9....

The 2005 AFRL/HEC One-Speaker Detection Systems

“…The training conditions all involved four-wire (two-channel) conversations and were defined by the following amounts of data: (1) an excerpt esti- 'The AFRLI/EC system submitted for the conditions requiring mated to contain approximately 10 seconds of speech of speaker segmentation and clustering is described in [2]. the target on its designated side (designated as 10sec4w), The GMM-based systems, regardless of feature set, all (designated as 10sec4w) or (2) one five-minute converused Version 2.1 of the MIT Lincoln Laboratory (MITsation (designated as lconv4w) as in the 10sec4w and LL) MFCC/GMM system [5] with 2048 mixtures per I conv4w training conditions, respectively, model and diagonal covariance matrices for each mixture.…”

“…the target on its designated side (designated as 10sec4w), The GMM-based systems, regardless of feature set, all (designated as 10sec4w) or (2) one five-minute converused Version 2.1 of the MIT Lincoln Laboratory (MITsation (designated as lconv4w) as in the 10sec4w and LL) MFCC/GMM system [5] with 2048 mixtures per I conv4w training conditions, respectively, model and diagonal covariance matrices for each mixture. In addition to the speech files, NIST provided tranAll of the GMM-based systems used a common scripts produced by an English-language speech recogspeech activity detector (SAD), 2 which worked in three nition system from BBN with word error rates typically stages. The first stage utilized a two-state speech/nonin the range of 15-30% for English conversational telespeech Hidden Markov Model (HMM) with MFCCs as phone speech.…”

Speaker diarization of news broacasts and meeting recordings

“…Application examples of the above include telephony mixed-channel speaker verification [Ore et al, 2006;Deng et al, 2006] and acoustic model adaptation for speech recognition [Pusateri & Hazen, 2002;Hain et al, 2006;Janin et al, 2006]. In the mixed-channel speaker verification task, speaker specific models after training are used to perform verification against a designated reference model.…”

“…Segmentation is performed at these locations and further speaker clustering can then be done so as to determine the identity of the speakers present. This is the strategy that was used in papers such as [Siu et al, 1992;Wegmann et al, 1999b;Kemp et al, 2000;Ore et al, 2006]. In [Wegmann et al, 1999b], an amplitude based silence detector is used as a first pass to break up continuous broadcast news recordings into segments.…”

“…In [Gish et al, 1991], speaker segmentation and clustering was performed on audio recordings of the radio dialogs between airport traffic controllers and pilots. Papers such as [Ore et al, 2006;Deng et al, 2006] have addressed performing diarization of mixed-channel telephone conversations as a first step to speaker verification in the NIST 2006 Speaker Recognition evaluations [NIST, 2006b]. Two of the most commonly researched domains is that of speaker diarization of broadcast news [Kubala et al, 1998;Cook & Robinson, 1998;Pallett et al, 1998] and meeting recordings [Fiscus et al, 2005;Anguera et al, 2005b;Hain et al, 2005].…”