Prosodic and other Long-Term Features for Speaker Diarization

Friedland, Gerald; Vinyals, Oriol; Huang, Yan; Müller, Christian

doi:10.1109/tasl.2009.2015089

Cited by 65 publications

(66 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…In other words, speaker clusters obtained with the bottom-up approach tend to be poorly normalized. This is particularly true when short-term cepstral-based features are used, though recent work with prosodic features have potential to discourage such behavior [10].…”

Section: Normalization and Discriminationmentioning

confidence: 99%

Linguistic influences on bottom-up and top-down clustering for speaker diarization

Bozonnet

Wang

Evans

et al. 2011

2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

While bottom-up approaches have emerged as the standard, default approach to clustering for speaker diarization we have always found the top-down approach gives equivalent or superior performance. Our recent work shows that significant gains in performance can be obtained when cluster purification is applied to the output of topdown systems but that it can degrade performance when applied to the output of bottom-up systems. This paper demonstrates that these observations can be accounted for by factors unrelated to the speaker and that they can impact more strongly on the performance of bottom-up clustering strategies than top-down strategies. Experimental results confirm that clusters produced through top-down clustering are better normalized against phone variation than those produced through bottom-up clustering and that this accounts for the observed inconsistencies in purification performance. The work highlights the need for marginalization strategies which should encourage convergence toward different speakers rather than toward nuisance factors such as that those related to the linguistic content.

show abstract

Section: Normalization and Discriminationmentioning

confidence: 99%

Linguistic influences on bottom-up and top-down clustering for speaker diarization

Bozonnet

Wang

Evans

et al. 2011

2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

show abstract

“…A dynamic programming procedure is used to find the optimal one-to-one mapping between the hypothesis and the ground truth segments so that the total overlap between the reference Category Feature ID Short description pitch f0 median median of the pitch pitch f0 min min of the pitch pitch f0 mean curve mean of the pitch tier formants f4 stddev std dev of the 4th formant formants f4 min min of the 4th formant formants f4 mean mean of the 4th formant formants f5 stddev std dev of the 5th formant formants f5 min min of the 5th formant formants f5 mean mean of the 5th formant harmonic harm mean mean of the harmonicsto-noise ratio formant form disp mean mean of the formant dispersion pitch pp period mean mean of the pointprocess of the periodicity contour Table 1. The 12 prosodic features used in the proposed initialization method (see also [2]). …”

Section: Baseline Systemmentioning

confidence: 99%

“…In this section, we present another method to estimate k (see Figure 1) and propose to use the aforementioned linear regression to adapt g accordingly. The presented method estimates the number of initial clusters and also provides a non-uniform initialization for the agglomerative clustering procedure based on the long-term feature study and ranking presented in [2]. Derived from the ranking in [2], the 12 topranked prosodic features (listed in Table 1) are extracted on all the speech regions (speech/non-speech detector, see [9]) in the recording.…”

Section: Automatic Parameter Estimation and Non-uniform Initializationmentioning

confidence: 99%

“…The presented method estimates the number of initial clusters and also provides a non-uniform initialization for the agglomerative clustering procedure based on the long-term feature study and ranking presented in [2]. Derived from the ranking in [2], the 12 topranked prosodic features (listed in Table 1) are extracted on all the speech regions (speech/non-speech detector, see [9]) in the recording. The features are extracted with the help of praatlib, a library that uses Praat 2 , on all the speech regions of the recordings (one-second-Hamming-window).…”

Section: Automatic Parameter Estimation and Non-uniform Initializationmentioning

confidence: 99%

“…We present how to use prosodic features, that have a good speaker discrimination ability according to the ranking method proposed in [2], to estimate the other key initialization parameter and to perform a non-uniform initialization. A comprehensive evaluation on the 2006, 2007, and 2009 NIST Rich Transcription evaluation sets (also known as RT-06, RT-07, and RT-09, respectively) compares the proposed approach to the ASPG model and to the state-of-the-art baseline system, resulting in a performance gain (from 12% up to 67% relative) compared to the baseline.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

An adaptive initialization method for speaker Diarization based on prosodic features

Friedland

2010

2010 IEEE International Conference on Acoustics, Speech and Signal Processing

Self Cite

View full text Add to dashboard Cite

The following article presents a novel, adaptive initialization scheme that can be applied to most state-of-the-art Speaker Diarization algorithms, i.e. algorithms that use agglomerative hierarchical clustering with Bayesian Information Criterion (BIC) and Gaussian Mixture Models (GMMs) of framebased cepstral features (MFCCs). The initialization method is a combination of the recently proposed "adaptive seconds per Gaussian" (ASPG) method and a new pre-clustering and number of initial clusters estimation method based on prosodic features. The presented initialization method has two important advantages. First, the method requires no manual tuning and is robust against file length and speaker count variations. Second, the method outperforms our previously used initialization methods on all benchmark files that were presented in the 2006, 2007, and 2009 NIST Rich Transcription (RT) evaluations and results in a Diarization Error Rate (DER) improvement of up to 67% (relative).

show abstract