## Enhancements to Transformation-Based Speaker Adaptation: Principal Component and Inter-Class Maximum Likelihood Linear Regression (2000)

Citations: | 5 - 1 self |

### BibTeX

@TECHREPORT{Doh00enhancementsto,

author = {Sam-joo Doh},

title = {Enhancements to Transformation-Based Speaker Adaptation: Principal Component and Inter-Class Maximum Likelihood Linear Regression},

institution = {},

year = {2000}

}

### OpenURL

### Abstract

iii Abstract In this thesis we improve speech recognition accuracy by obtaining better estimation of linear transformation functions with a small amount of adaptation data in speaker adaptation. The major contributions of this thesis are the developments of two new adaptation algorithms to improve maximum likelihood linear regression. The first one is called principal component MLLR (PC-MLLR), and it reduces the variance of the estimate of the MLLR matrix using principal component analysis. The second one is called inter-class MLLR, and it utilizes relationships among different transformation functions to achieve more reliable estimates of MLLR parameters across multiple classes. The main idea of PC-MLLR is that if we estimate the MLLR matrix in the eigendomain, the variances of the components of the estimates are inversely proportional to their eigenvalues. Therefore we can select more reliable components to reduce the variances of the resulting estimates and to improve speech recognition accuracy. PC-MLLR eliminates highly variable components and chooses the principal components corresponding to the largest eigenvalues. If all the component are used, PC-MLLR becomes the same as conventional MLLR. Choosing fewer principal components increases the bias of the estimates which can reduce recognition accuracy. To compensate for this problem, we developed weighted principal component MLLR (WPC-MLLR). Instead of eliminating some of the components, all the components in WPC-MLLR are used after applying weights that minimize the mean square error. The component corresponding to a larger eigenvalue has a larger weight than the component corresponding to a smaller eigenvalue. As more adaptation data become available, the benefits from these methods may become smaller because ...

### Citations

3927 |
Pattern classification and scene analysis
- Duda, Hart
- 1973
(Show Context)
Citation Context ...stribution is known to have a Gaussian distribution λ MAP = argmaxλ( P( λ X) ) = argmaxλ( PXλ ( )P( λ) ) Pxμ ( ) N μσ 2 ∼ ( , ) 2 P( μ) ∼ N( μo, σo ) The MAP estimate of a mean μ is o=-=btained by (e.g. [15]) μˆ MAP = 2 N�-=-�o ------------------------- x 2 2 Nσo + σ = βx + ( 1 – β)μo σ 2 + -------------------------μ 2 2 o Nσo + σ N 2 1 Nσo where x = --- x is the sample mean, and is a weighting factor. N ∑ n... |

2026 |
Principal Component Analysis
- Jolliffe
- 1986
(Show Context)
Citation Context ...e full matrix in their test. This is because they already had enough adaptation data to estimate the full matrix, or because three blocks may not have been optimal. Principal component analysis (PCA) =-=[31]-=- has been used to reduce the dimensionality of the data. The original data which consisted of interrelated variables are transformed into a new set of uncorrelated variables by the eigenvectors of the... |

804 | Fundamentals of statistical signal processing: estimation theory - Kay |

740 |
Statistical methods for speech recognition
- Jelinek
- 1997
(Show Context)
Citation Context ...2 where bs( o) is an output probability for state s, wsk is a mixing weight, μsk is a mean vector, and Csk is a covariance matrix for k th Gaussian in state s. More details about HMMs can be found in=-= [24, 30, 47]-=-. S 2 Figure 2.2 An example of hidden Markov models (HMM). A 3-state model is depicted, which represents a phoneme. Models are concatenated to represent a word or sentence. S i = State i p ij = Transi... |

593 | Maximum likelihood linear regression for speaker adaptation of continuous density hidden markov models
- Leggetter, Woodland
- 1995
(Show Context)
Citation Context ... transformation parameters. One of the most successful implementations of transformation-based adaptation is the Maximum Likelihood Linear Regression (MLLR) method developed by Leggetter and Woodland =-=[40].-=- MLLR applies a linear transformation on recognition model parameters (which are usually Gaussian mean vectors) to obtain an adapted model. If μk is the mean vector of a Gaussian mixture component in... |

491 | Maximum a posteriori estimation for multivariate Gaussian mixture observations of Markov chains
- GAUVAIN, C-H
- 1994
(Show Context)
Citation Context ...ormation. An early seminal study on MAP adaptation for speech recognition by Lee et al. [37] described how to reestimate the mean and covariance parameters of continuous density HMMs. Gauvain and Lee =-=[18]-=- further extended this work to estimate other HMM parameters. The Bayesian approach has been applied to semi-continuous HMMs and discrete density HMMs by Huo et al. [26], among many others. 2.3.2 Tran... |

276 | Hidden Markov models for speech recognition
- Huang, Ariki, et al.
- 1990
(Show Context)
Citation Context ...2 where bs( o) is an output probability for state s, wsk is a mixing weight, μsk is a mean vector, and Csk is a covariance matrix for k th Gaussian in state s. More details about HMMs can be found in=-= [24, 30, 47]-=-. S 2 Figure 2.2 An example of hidden Markov models (HMM). A 3-state model is depicted, which represents a phoneme. Models are concatenated to represent a word or sentence. S i = State i p ij = Transi... |

267 |
Applied Linear Statistical Models
- Neter, Kutner, et al.
- 1996
(Show Context)
Citation Context ...ompare speaker adaptation performance obtained using the methods described. 4.2 Principal Component Regression Principal component regression was developed in classical linear regression theory (e.g. =-=[42, 43]-=-). For example, consider N pairs of samples (xi , yi ), where xi is a D-dimensional row vector, and yi and exi are 2 scalars. exi is assumed to have a Gaussian distribution with zero mean and variance... |

237 |
RH: Classical and Modern Regression with Applications. 2nd edition
- Myers
- 1990
(Show Context)
Citation Context ...ompare speaker adaptation performance obtained using the methods described. 4.2 Principal Component Regression Principal component regression was developed in classical linear regression theory (e.g. =-=[42, 43]-=-). For example, consider N pairs of samples (xi , yi ), where xi is a D-dimensional row vector, and yi and exi are 2 scalars. exi is assumed to have a Gaussian distribution with zero mean and variance... |

197 | Some statistical issues in the comparison of speech recognition algorithms
- Gillick, Cox
- 1989
(Show Context)
Citation Context ...er has 5.55 utterances on average, and each utterance is about 2-4 seconds long. Adaptation is done in unsupervised mode. In this task, the statistical significance test using tools developed at NIST =-=[19]-=- indicates that if two algorithms produce 0.2% absolute or larger difference in WER then they are statistically different [25]. Even though the improvements after adaptation is small, WPC-MLLR provide... |

169 | Environmental robustness in automatic speech recognition
- Acero, Stem
- 1990
(Show Context)
Citation Context ...input feature vectors or the recognition model parameters as shown in Fig. 2.4 [36]. Cepstral mean normalization (CMN) (or cepstral mean subtraction [48]), codedependent cepstral normalization (CDCN) =-=[1]-=-, and vocal tract length normalization [20] are some of the examples of the method which modifies input feature vectors. Sankar and Lee [51] proposed a maximumlikelihood stochastic matching approach t... |

106 |
Regression Analysis by Example
- CHATTERJEE, HADI, et al.
- 2000
(Show Context)
Citation Context ...x i – bm) i = 1 2σe 2 N = = ⋅ exp ∑ i = 1 Q m ----- 1 ( y 2 i – amx i – bm) σe 2 N = ∑ i = 1 y = a x + b 2 We can see that Eqs. (3.3) and (3.5) give the same result if σe is constant. T=-=he solution is [5] x (-=-3.4) (3.5)sChapter 3. Maximum Likelihood Linear Regression 26 N N where x --- 1 1 = x is the sample mean of , and is the sample mean of . N ∑ i xi y = --- N ∑ yi yi i = 1 i = 1 2 If ei has a diffe... |

106 | A Maximum-Likelihood Approach to Stochastic Matching for Robust Speech Recognition
- Sankar, Lee
(Show Context)
Citation Context ...btraction [48]), codedependent cepstral normalization (CDCN) [1], and vocal tract length normalization [20] are some of the examples of the method which modifies input feature vectors. Sankar and Lee =-=[51]-=- proposed a maximumlikelihood stochastic matching approach to estimate an inverse distortion function. These methods modify all input feature vectors, which may require too much computation. On the co... |

75 | Structural maximum a posteriori linear regression for fast HMM adaptation
- Siohan, Myrvoll, et al.
- 2002
(Show Context)
Citation Context ...2.3)sChapter 2. Background Information 17 However, we can apply MAP criteria to a transformation function too. For example, the MAP criterion has been applied to linear regression in the MAPLR method =-=[7]-=-. In transformation-based adaptation, we usually assume that the number of transformation parameters is much smaller than the target parameters. However, we can increase the number of transformation f... |

62 | The generation and use of regression class trees for MLLR adaptation
- Gales
- 1996
(Show Context)
Citation Context ...Class 1 f 2( μk) , k ∈ Class 2 Method II f ( . ) f ( . ) 1 2 Class 1 Class 2 Figure 5.1 Classification of Gaussian mean vectors as either a single class (Method I) or as two classes (Method II) Gal=-=es [16]-=- used regression class trees to find the optimal number of regression classes in conventional MLLR. He developed a binary regression class tree and dynamically selected the classes so that either each... |

61 |
A Study on Speaker Adaptation of the Parameters of Continuous Density Hidden Markov Models
- Lee, Lin, et al.
- 1991
(Show Context)
Citation Context ...e prior”) then the MAP estimate is approximately equal to the ML estimate because we don’t have useful prior information. An early seminal study on MAP adaptation for speech recognition by Lee et =-=al. [37]-=- described how to reestimate the mean and covariance parameters of continuous density HMMs. Gauvain and Lee [18] further extended this work to estimate other HMM parameters. The Bayesian approach has ... |

47 | Speaker Adaptation Using Combined Transformation and Bayesian Methods
- Digalakis, Neumeyer
- 1996
(Show Context)
Citation Context ...hared variance scaling are assumed, and estimated using the MAP framework. Other research groups have proposed hybrid algorithms to combine the transformation-based and Bayesian adaptation approaches =-=[8, 11]-=-. Shinoda and Lee [52] proposed a structured MAP (SMAP) adaptation, in which the transformation parameters are estimated in a hierarchical structure. Despite its name, the SMAP algorithm does not make... |

44 | Variance compensation within the MLLR framework
- Gales, Woodland
- 1996
(Show Context)
Citation Context ...s of the test data, resulting in low recognition accuracy. Therefore we would like to obtain more reliable estimates of A and b. We may achieve this by reducing the number of parameters. Gales et al. =-=[17]-=- constrained the transformation matrix A to a block diagonal structure, with feature components assumed to have correlation only within a block. This reduces the number of parameters to be estimated. ... |

44 |
Dynamic Programming Optimization for Spoken Word Recognition
- Sakoe, Chiba
- 1978
(Show Context)
Citation Context ... changes every time it is spoken (temporal variation). Pattern Decision Output Comparison Rule Text Reference Patterns Dynamic Time Warping (DTW) has been developed to overcome the temporal variation =-=[50]. It-=- stores a set of reference patterns, and aligns test and reference patterns so that they their smallest “distance” between them according to a predetermined metric. After comparing the test word w... |

39 |
Subphonetic Acoustic Modeling for Speaker-Independent Continuous Speech Recognition
- Hwang
- 1993
(Show Context)
Citation Context ... the recognition units to consider the influence by the neighboring phonemens. Because the number of states of HMM for the triphone is too large, tied states, or senones are used to reduce the number =-=[29]. -=-The output distribution of a state is represented by the Gaussian mixtures. 6000 senones and 16 Gaussians per senone are used for Wall Street Journal task in this thesis. 2.4.2 Test Corpora • Wall S... |

38 |
On stochastic feature and model compensation approaches to robust speech recognition
- Lee
- 1998
(Show Context)
Citation Context ..., w1) i = 1sChapter 2. Background Information 9 2.3 Review of Speaker Adaptation For adaptation we can either modify the input feature vectors or the recognition model parameters as shown in Fig. 2.4 =-=[36]-=-. Cepstral mean normalization (CMN) (or cepstral mean subtraction [48]), codedependent cepstral normalization (CDCN) [1], and vocal tract length normalization [20] are some of the examples of the meth... |

36 |
Signal Bias Removal by Maximum Likelihood Estimation for Robust Telephone Speech Recognition
- Rahim, Juang
- 1996
(Show Context)
Citation Context ...Adaptation For adaptation we can either modify the input feature vectors or the recognition model parameters as shown in Fig. 2.4 [36]. Cepstral mean normalization (CMN) (or cepstral mean subtraction =-=[48]-=-), codedependent cepstral normalization (CDCN) [1], and vocal tract length normalization [20] are some of the examples of the method which modifies input feature vectors. Sankar and Lee [51] proposed ... |

33 |
Improved acoustic Modeling for HMMs using Linear Transformations
- Leggetter
- 1995
(Show Context)
Citation Context ...-MLLR . . . . 53 Table 5.1 The actual values used in the simulation in Fig. 5.3 . . . . . . . . . . . . . . . . 61 Table 5.2 Phonetic-based regression classes used in inter-class MLLR (from Leggetter =-=[39]-=- with minor modifications) . . . . . . . . . . . . . . . . . . . . . . 67 Table 5.3 Word error rates for non-native speakers (s3-94) after supervised adaptation using inter-class MLLR (Relative percen... |

26 | Bayesian adaptive learning of the parameters of hidden Markov model for speech recognition
- Huo, Chan, et al.
- 1995
(Show Context)
Citation Context ... density HMMs. Gauvain and Lee [18] further extended this work to estimate other HMM parameters. The Bayesian approach has been applied to semi-continuous HMMs and discrete density HMMs by Huo et al. =-=[26]-=-, among many others. 2.3.2 Transformation-Based Adaptation In transformation-based adaptation, we assume that a set of the target parameters is updated by the same transformation. When we try to adapt... |

21 | On-line adaptive learning of the correlated continuous density hidden Markov models for speech recognition
- Huo, Lee
- 1998
(Show Context)
Citation Context ...nversion of a large matrix. Rozzi and Stern [49] developed a least mean square (LMS) algorithm for an efficient computation of the EMAP algorithm at the expense of a finite misadjustment. Huo and Lee =-=[27]-=- proposed an algorithm based on pairwise correlation of parameter pairs to reduce computational complexity. Afify et al [3] also used pairwise correlation of parameters. For a target parameter, they g... |

19 |
Combined Bayesian and predictive techniques for rapid speaker adaptation of continuous density hidden Markov models
- Ahadi, Woodland
- 1997
(Show Context)
Citation Context ...tor is estimated by interpolating the neighboring vectors, and trained vectors are smoothed to reduce estimation errors. Cox [10] predicted unheard sounds using correlation models. Ahadi and Woodland =-=[2]-=- proposed the regression-based model prediction (RMP) technique. They applied linear regression to estimate parametric relationships among the model parameters, and used it to update those parameters ... |

19 |
Vector-fieldsmoothed Bayesian Learning for Incremental Speaker Adaptation
- Takahashi, Sagayama
- 1995
(Show Context)
Citation Context ...ces MLLR when they applied ABC with MLLR on the F0 condition of the ARPA HUB 4-96 evaluation data (which deals with the recognition of clean prepared speech from broadcast news shows). Sagayama et al =-=[53, 54]-=- suggested the Vector Field Smoothing (VFS) method to overcome insufficient training data. An unobserved vector is estimated by interpolating the neighboring vectors, and trained vectors are smoothed ... |

18 |
A posteriori estimation of correlated jointly Gaussian mean vectors
- Lasry, Stern
- 1984
(Show Context)
Citation Context ...f there are no adaptation data associated with it. Several methods have been proposed to overcome this disadvantage, which mostly utilize the relationships among different parameters. Lasry and Stern =-=[35]-=- introduced the extended MAP (EMAP) adaptation which makes use of information about correlations among parameters. A parameter can be adapted even though there is no adaptation data directly associate... |

17 |
Predictive Speaker Adaptation in Speech Recognition
- Cox
- 1995
(Show Context)
Citation Context ...hing (VFS) method to overcome insufficient training data. An unobserved vector is estimated by interpolating the neighboring vectors, and trained vectors are smoothed to reduce estimation errors. Cox =-=[10]-=- predicted unheard sounds using correlation models. Ahadi and Woodland [2] proposed the regression-based model prediction (RMP) technique. They applied linear regression to estimate parametric relatio... |

17 |
Unsupervised Adaptation Using Structural Bayes Approach
- Shinoda, Lee
- 1998
(Show Context)
Citation Context ...e assumed, and estimated using the MAP framework. Other research groups have proposed hybrid algorithms to combine the transformation-based and Bayesian adaptation approaches [8, 11]. Shinoda and Lee =-=[52]-=- proposed a structured MAP (SMAP) adaptation, in which the transformation parameters are estimated in a hierarchical structure. Despite its name, the SMAP algorithm does not make use of prior informat... |

16 |
A Bayesian predictive classification approach to robust speech recognition
- Huo, Jiang, et al.
- 1997
(Show Context)
Citation Context ...rocedure we will need to estimate or predict the characteristics of the test speaker. Similar topics have been studied in Bayesian framework where the prior distributions of model parameters are used =-=[28]-=-. 6.2.2 Control of Weights in Inter-Class MLLR Using neighboring classes to estimate the parameters in the target class has both advantages and disadvantages. Neighboring classes are helpful because t... |

12 | Speech Recognition in Mobile Environments
- HUERTA
(Show Context)
Citation Context ...entences each. The Spoke 0 is similar to the Spoke 3 except that they are from native speakers. • The Telefónica (TID) corpus This corpus was recorded over the GSM cellular telephone network in Spa=-=in [25]-=-. It is a small vocabulary set (approximately 75 words in lexicon) which contain mainly numbers. It includes approximately 6000 training utterances and 3000 testing utterances (with approximately 1450... |

10 | Tied-structure HMM based on parameter correlation for efficient model training - Takahashi, Sagayama - 1996 |

10 |
Maximum a Posteriori Adaptation Techniques for Speech Recognition
- Zavaliagkos
- 1995
(Show Context)
Citation Context ...d MAP (EMAP) adaptation which makes use of information about correlations among parameters. A parameter can be adapted even though there is no adaptation data directly associated with it. Zavaliagkos =-=[56]-=- was the first to apply EMAP adaptation to large scale HMM speech recognizers. While the EMAP approach produces an elegant analytical adaptation equation that makes appropriate use of correlations amo... |

8 |
Correlation based predictive adaptation of hidden Markov models
- Afify, Gong, et al.
- 1997
(Show Context)
Citation Context ...of the EMAP algorithm at the expense of a finite misadjustment. Huo and Lee [27] proposed an algorithm based on pairwise correlation of parameter pairs to reduce computational complexity. Afify et al =-=[3]-=- also used pairwise correlation of parameters. For a target parameter, they got several estimates using correlations with different parameters, and combine them to get the final estimate for the targe... |

8 | Speaker adaptation by correlation (ABC
- Chen, DeSouza
- 1997
(Show Context)
Citation Context ...n of parameters. For a target parameter, they got several estimates using correlations with different parameters, and combine them to get the final estimate for the target parameter. Chen and DeSouza =-=[6]-=- also used the correlation between speech units for a similar speaker adaptation algorithm which they refer to as Adaptation By Correlation (ABC). The estimates are derived using least squares theory,... |

8 |
A hybrid algorithm for speaker adaptation using MAP transformation and adaptation
- Chien, Lee, et al.
- 1997
(Show Context)
Citation Context ...hared variance scaling are assumed, and estimated using the MAP framework. Other research groups have proposed hybrid algorithms to combine the transformation-based and Bayesian adaptation approaches =-=[8, 11]-=-. Shinoda and Lee [52] proposed a structured MAP (SMAP) adaptation, in which the transformation parameters are estimated in a hierarchical structure. Despite its name, the SMAP algorithm does not make... |

8 |
Feature selection methods for hidden markov model-based speech recognition
- Nouza
- 1996
(Show Context)
Citation Context ...data. The original data which consisted of interrelated variables are transformed into a new set of uncorrelated variables by the eigenvectors of the covariance matrix of the original data set. Nouza =-=[44] use-=-d PCA for feature selection in a speech recognition system. Kuhn et al. [34] introduced “eigenvoices” to represent the prior knowledge of speaker variation. Hu [23] applied PCA to describe the cor... |

6 | Inter-Class MLLR for Speaker Adaptation
- Doh, Stern
- 2000
(Show Context)
Citation Context ...quare Error 0.2 0.15 0.1 0.05 5.3 Inter-Class MLLR We apply the method described in the previous section to the MLLR framework, and develop a new adaptation procedure which is called inter-class MLLR =-=[13]-=-. In inter-class MLLR the inter-class transformation is given by another linear regression. The inter-class transformation provides a priori knowledge among the regression classes, and helps to estima... |

6 | Vector Field Smoothing Principle for Speaker Adaptation - Hattori, Sagayama - 1992 |

6 |
et al., “Eigenvoices for speaker adaptation
- Kuhn
- 1998
(Show Context)
Citation Context ...d into a new set of uncorrelated variables by the eigenvectors of the covariance matrix of the original data set. Nouza [44] used PCA for feature selection in a speech recognition system. Kuhn et al. =-=[34] int-=-roduced “eigenvoices” to represent the prior knowledge of speaker variation. Hu [23] applied PCA to describe the correlation between phoneme classes for speaker normalization. While the general mo... |

5 | Improved Estimation of Supervision in Unsupervised Speaker Adaptation - Homma, Aikawa, et al. |

5 |
Understanding and Adapting to Speaker Variability Using Correlation-Based Principal Analysis
- Hu
- 1999
(Show Context)
Citation Context ...of the original data set. Nouza [44] used PCA for feature selection in a speech recognition system. Kuhn et al. [34] introduced “eigenvoices” to represent the prior knowledge of speaker variation.=-= Hu [23]-=- applied PCA to describe the correlation between phoneme classes for speaker normalization. While the general motivation for these approaches was similar to the approach described in this thesis, none... |

5 | Smoothed N-best-based speaker adaptation for speech recognition - Matsui, Matsuoka, et al. - 1997 |

4 |
Extraction of Reliable Transformation Parameters for Unsupervised Speaker Adaptation
- Chien, Junqua, et al.
- 1999
(Show Context)
Citation Context ...imized. He also described (5.1) (5.2) (5.3)sChapter 5. Inter-Class MLLR 57 a method in which MLLR was applied to the root node and the sub-nodes of the regression class tree iteratively. Chien et al. =-=[9]-=- described a hierarchical adaptation procedure that used a tree structure to control the transformation sharing and to obtain reliable transformation parameters. As we have noted above, new parameters... |

4 | Speaker adaptation in continuous speech recognition via estimation of correlated mean vectors
- Rozzi, Stern
- 1991
(Show Context)
Citation Context ... an elegant analytical adaptation equation that makes appropriate use of correlations among adaptation parameters, solution of the equation depends on the inversion of a large matrix. Rozzi and Stern =-=[49]-=- developed a least mean square (LMS) algorithm for an efficient computation of the EMAP algorithm at the expense of a finite misadjustment. Huo and Lee [27] proposed an algorithm based on pairwise cor... |

3 | Weighted principal component MLLR for speaker adaptation
- Doh, Stern
- 1999
(Show Context)
Citation Context ...e larger. We determine the best p value in advance using development data which are similar to the actual test data. 4.3 Principal Component MLLR The formulation of Principal Component MLLR (PC-MLLR) =-=[12] is -=-very similar to that discussed in the previous section, except that we also consider the a posteriori probability γ t( k) as well as the baseline 2 variance σk and the shift vector b . Let’s consi... |

2 | Acoustic-Feature-Based Frequency Warping for Speaker Normalization
- Gouvea
- 1998
(Show Context)
Citation Context ...odel parameters as shown in Fig. 2.4 [36]. Cepstral mean normalization (CMN) (or cepstral mean subtraction [48]), codedependent cepstral normalization (CDCN) [1], and vocal tract length normalization =-=[20]-=- are some of the examples of the method which modifies input feature vectors. Sankar and Lee [51] proposed a maximumlikelihood stochastic matching approach to estimate an inverse distortion function. ... |

2 |
et al.: 1994 Benchmark Tests for the ARPA Spoken Language Program
- Pallett, Fiscus, et al.
(Show Context)
Citation Context ...especially on the non-native English speakers from Spoke 3 in the 1994 DARPA evaluation (s3-94). Spoke 3 is designed “to evaluate a rapid enrollment speaker adaptation algorithm on difficult speaker=-=s [45].�-=-�� It consists of 10 speakers reading 20 test sentences each. Each speaker reads news articles from the WSJ corpus. The primary test in Spoke 3 uses 40 adaptation sentences for each speaker in supervi... |

2 |
et al, “1998 Broadcast news benchmark test results: English and non-English word error rate performance measures
- Pallett
(Show Context)
Citation Context ...onditions. It contains a broad sample of speaker and environmental conditions. • Broadcast News from the 1998 DARPA Evaluation (Hub 4) The Hub 4 corpus contains recorded speech from TV broadcast new=-=s [46]-=-. It is a mixture of several different conditions, e.g. planned/spontaneous, clean/noisy, full/narrow bandwidth, and native/nonnative. It has a much larger vocabulary than the previous corpora. The ba... |