重叠语音是影响说话人分割性能的主要因素之一。该文提出了基于语音高层信息特征的重叠语音检测方法以提高说话人分割效果。首先用通用背景模型（universal background model，UBM）提取语音的语言学高层信息特征，并融合这些特征和Mel频率倒谱系数（Mel frequency cepstral coefficient，MFCC）特征建立隐Markov模型（hidden Markov model，HMM）检测重叠语音，然后对处理后的语音进行说话人分割。实验结果表明：对于由TIMIT语音库生成的数据集，该方法对重叠语音检测的错误率比单一采用MFCC特征有显著降低，而且说话人分割性能有明显的提高。
Overlapping speech is one of the main factors influencing the performance of speaker segmentation. This paper presents an overlapping speech detection method using a high-level information feature to improve the speaker segmentation results. A linguistic high-level information feature of the speech is extracted using the universal background model (UBM). Then, a hidden Markov model (HMM) is trained using the Mel frequency cepstral coefficients (MFCC) and the high-level information to detect overlapping speech. The result is then used for the speaker segmentation of the pre-processed speech. Tests on a dataset generated from the TIMIT database show that the error ratio for overlapping speech detection is significantly lower than the reference method using just the MFCC feature. The speaker segmentation is also significantly improved.
[1] Shriberg E, Stolcker A, Baron D. Observations on overlap:Finding and implications for automatic processing of multi-party conversation[C]//Proc 7th European Conference on Speech Communication and Technology. Aalborg, Denmark:ISCA, 2001:1359-1362. [2] Sinclair M, King S. Where are the challenges in speaker diarization[C]//Proc International Conference on Acoustics, Speech, Signal and Signal Processing. Vancouver, Canada:IEEE, 2013:7741-7745. [3] 马勇, 鲍长春. 说话人分割聚类研究进展[J]. 信号处理, 2013, 29(9):1190-1199.MA Yong, BAO Changchun. Advance in speaker segmentation and clustering[J]. Journal of Signal Processing, 2013, 29(9):1190-1199. (in Chinese). [4] Kotti M, Moschou V, Kotropoulos C. Speaker segmentation and clustering[J]. Signal Processing, 2008. 88(5):1091-1124. [5] Otterson S, Ostendorf M. Efficient use of overlap information in speaker diarization[C]//Proc Conference Automatic Speech Recognition & Understanding, Kyoto, Japan:IEEE, 2007:683-686. [6] Roakye K, Hornero B, Vinyals O, et al. Overlapped speech detection for improved diarization in multi-party meetings[C]//Proc International Conference on Acoustics, Speech, Signal and Signal Processing. Las Vegas, NV, USA:IEEE, 2008:4353-4356. [7] Roakye K, Vinyals O, Friedland G. Improved overlapped speech handling for speaker diarization[C]//Proc International Speech Communication Association. Florence, Italy:ISCA, 2011:941-944. [8] Zelenak M, Segura C, Luque J, et al, Simultaneous speech detection with spatial features for speaker diarization[J]. IEEE Transaction on Audio, Speech and Language Processing, 2012, 20(2):436-446. [9] Geiger J T, Eyben F, Evans N, et al. Using linguistic information to detect overlapping speech[C]//Proc International Speech Communication Association. Lyon, France:ISCA, 2013:941-944. [10] Yella S H, Bourlard H. Overlapping speech detection using long-term conversational features for speaker diarization in meeting room conversations[J]. IEEE Transaction on Audio, Speech and Language Processing, 2014, 22(12):1688-1700. [11] Reynolds D, Quatieri T, Dunn R. Speaker verification using adapted Gaussian mixture models[J]. Digital Signal Processing, 2000, 10(1):19-41. [12] Delacourt P, Wellekens C J. DISTBIC:A speaker-based segmentation for audio data indexing[J]. Speech Communication, 2000, 32(1):111-126.
