Boosted MMI for model and feature-space discriminative training

Povey, Daniel; Kanevsky, Dimitri; Kingsbury, Brian; Ramabhadran, Bhuvana; Saon, George; Visweswariah, Karthik

doi:10.1109/icassp.2008.4518545

Cited by 278 publications

(176 citation statements)

References 7 publications

Supporting

Mentioning

175

Contrasting

Unclassified

Order By: Relevance

“…-Thông tin tương hỗ cực đại tăng cường (BMMI -Boosted Maximum Mutual Information): sử dụng biến thể của hàm giá MMI với hệ số tăng cường để làm tăng các mô hình có thể trộn được [15].…”

Section: Mô Hình âM Họcunclassified

Cải Thiện Hiệu Năng Hệ Thống Nhận Dạng Tiếng Việt Với Thông Tin Về Phương Ngữ

Hùng¹,

Loan²,

Quang³

et al. 2017

Fair - Nghiên Cứu Cơ Bản Và Ứng Dụng Công Nghệ Thông Tin - 2016

View full text Add to dashboard Cite

Section: Mô Hình âM Họcunclassified

Cải Thiện Hiệu Năng Hệ Thống Nhận Dạng Tiếng Việt Với Thông Tin Về Phương Ngữ

Hùng¹,

Loan²,

Quang³

et al. 2017

Fair - Nghiên Cứu Cơ Bản Và Ứng Dụng Công Nghệ Thông Tin - 2016

View full text Add to dashboard Cite

“…Unfortunately, there is no natural choice of accuracy. Two convenient definitions are the approximate accuracy also used for MPE [1,8] and the framebased accuracy (Hamming distance) [4,9]. In the remainder of this section, two variants of the above smooth optimization criterion are introduced and discussed.…”

Section: Modified Training Criteria In Asrmentioning

confidence: 99%

“…Due to the equivalence of GHMMs and Gaussian-like HCRFs [6], we can also use GHMMs for the acoustic model and i-smoothing for regularization [1]. Note that it can be shown that this variant of Modified MMI is theoretically equivalent to the heuristically motivated Boosted MMI [8]. In practice, Boosted MMI differs from Modified MMI in the way the iteration constants are set and the choice of the acoustic model for i-smoothing.…”

Section: Modified MMImentioning

confidence: 99%

Modified MPE/MMI in a transducer-based framework

Heigold

Schlüter

Ney

2009

2009 IEEE International Conference on Acoustics, Speech and Signal Processing

View full text Add to dashboard Cite

In this paper we show how common training criteria like for example MPE or MMI can be extended to incorporate a margin term. In addition, a transducer-based training implementation is presented, which covers a large variety of discriminative training criteria for ASR, including the standard MMI, MPE, and MCE criteria, as well as the modifications to these criteria presented here. The modified criteria are directly related with the conventional large margin formulation of SVMs. In the proposed approach, we can take advantage of the generalization guarantees of large margin classifiers while keeping the existing framework for the discriminative training, including the efficient algorithms for conventional MPE or MMI. On the conceptual side, this allows for a direct evaluation of the margin term. Finally, experimental results are presented for different large vocabulary continuous speech recognition tasks (one of which is trained on a very large amount of training data) using these modified criteria.

show abstract

“…As explained in [17], the accuracy is generally scaled with some ρ > 0, and this term weighs up the likelihoods of the competing hypotheses compared with the correct hypothesis [36]. On the contrary, this term can be equally interpreted as a margin term.…”

Section: Margin-based Maximum Mutual Informationmentioning

confidence: 99%

Confidence- and margin-based MMI/MPE discriminative training for off-line handwriting recognition

Dreuw

Heigold

Ney

2011

IJDAR

View full text Add to dashboard Cite

We present a novel confidence-and marginbased discriminative training approach for model adaptation of a hidden Markov model (HMM) based handwriting recognition system to handle different handwriting styles and their variations.Most current approaches are maximum-likelihood (ML) trained HMM systems and try to adapt their models to different writing styles using writer adaptive training, unsupervised clustering, or additional writer specific data. Here, discriminative training based on the maximum mutual information (MMI) and minimum phone error (MPE) criteria are used to train writer independent handwriting models. For model adaptation during decoding, an unsupervised confidence-based discriminative training on a word and frame level within a two-pass decoding process is proposed.The proposed methods are evaluated for closedvocabulary isolated handwritten word recognition on the IFN/ENIT Arabic handwriting database, where the word-error-rate is decreased by 33% relative compared to a ML trained baseline system. On the largevocabulary line recognition task of the IAM English handwriting database, the word-error-rate is decreased by 25% relative.

show abstract

Boosted MMI for model and feature-space discriminative training

Cited by 278 publications

References 7 publications

Cải Thiện Hiệu Năng Hệ Thống Nhận Dạng Tiếng Việt Với Thông Tin Về Phương Ngữ

Cải Thiện Hiệu Năng Hệ Thống Nhận Dạng Tiếng Việt Với Thông Tin Về Phương Ngữ

Modified MPE/MMI in a transducer-based framework

Confidence- and margin-based MMI/MPE discriminative training for off-line handwriting recognition

Contact Info

Product

Resources

About