On the phonetic structure of a large hidden Markov model

Pepper, D.J.; Clements, Mark A.

doi:10.1109/icassp.1991.150377

Cited by 8 publications

(4 citation statements)

References 4 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The results we obtained can be compared with other recent reports [5], [14], [15]. Bocchieri and Wilpon [13] employed 38-dimensional delta-delta cepstrum (DDCEP) feature vectors and obtained 98.6% recognition accuracy for the TI connected digits and 80.8% for the E-set letters.…”

Section: Resultssupporting

confidence: 62%

“…De Haan and Ececioglu reported 97.2% recognition accuracy for the TI isolated digits by using a 19-dimensional feature vector and a feature map strategy [5]. Another interesting result is 50.6% recognition accuracy for TIMIT phoneme recognition reported by Pepper and Clements, who employed 25-dimensional feature vectors and a large 128-state HMM [14].…”

Section: Resultsmentioning

confidence: 93%

See 1 more Smart Citation

Phoneme-based vector quantization in a discrete HMM speech recognizer

Zhang¹,

Togneri

Alder

1997

IEEE Trans. Speech Audio Process.

View full text Add to dashboard Cite

Section: Resultssupporting

confidence: 62%

Section: Resultsmentioning

confidence: 93%

Phoneme-based vector quantization in a discrete HMM speech recognizer

Zhang¹,

Togneri

Alder

1997

IEEE Trans. Speech Audio Process.

View full text Add to dashboard Cite

“…While no direct comparison with the previously established HMM segment vocoders of [T98] or [MTK98] is done, this method shows that an ergodic HMM with number of states as 128 yields a good overall quality and intelligibility with speaker characteristics preserved at an effective bit-rate of 128 bps, though no formal listening tests are done. The notions of the optimal number of states being 64 for American English to correspond to the phones in this work is based on earlier work [Pepp90,FC86,Pepp91] and is closely in corroboration with a related work on using ergodic-HMMs to model spoken languages for language-identification [RSS03, SR05].…”

Section: Ergodic Hmm Frameworkmentioning

confidence: 55%

Ultra Low Bit-Rate Speech Coding

Ramasubramanian¹,

Doddala²

2015

SpringerBriefs in Electrical and Computer Engineering

View full text Add to dashboard Cite

The authors of this series have been hand-selected. They comprise some of the most outstanding scientists -drawn from academia and private industry -whose research is marked by its novelty, applicability, and practicality in providing broad based speech solutions. The SpringerBriefs in Speech Technology series provides the latest findings in speech technology gleaned from comprehensive literature reviews and empirical investigations that are performed in both laboratory and real life settings. Some of the topics covered in this series include the presentation of real life commercial deployment of spoken dialog systems, contemporary methods of speech parameterization, developments in information security for automated speech, forensic speaker recognition, use of sophisticated speech analytics in call centers, and an exploration of new methods of soft computing for improving human-computer interaction. Those in academia, the private sector, the self-service industry, law enforcement, and government intelligence, are among the principal audience for this series, which is designed to serve as an important and essential reference guide for speech developers, system designers, speech engineers, linguists, and others. In particular, a major audience of readers will consist of researchers and technical experts in the automated call center industry where speech processing is a key component to the functioning of customer care contact centers.Amy Neustein, Ph.D., serves as Editor-in-Chief of the International Journal of Speech Technology (Springer). She edited the recently published book "Advances in Speech Recognition: Mobile Environments, Call Centers and Clinics" (Springer 2010), and serves as quest columnist on speech processing for Womensenews. Dr. Neustein is Founder and CEO of Linguistic Technology Systems, a NJ-based think tank for intelligent design of advanced natural language based emotion-detection software to improve human response in monitoring recorded conversations of terror suspects and helpline calls. Dr. Neustein's work appears in the peer review literature and in industry and mass media publications. Her academic books, which cover a range of political, social and legal topics, have been cited in the Chronicles of Higher Education, and have won her a Pro Humanitate Literary Award. She serves on the visiting faculty of the National Judicial College and as a plenary speaker at conferences in artificial intelligence and computing. Dr. Neustein is a member of MIR (machine intelligence research) Labs, which does advanced work in computer technology to assist underdeveloped countries in improving their ability to cope with famine, disease/illness, and political and social affliction. She is a founding member of the New York City Speech Processing Consortium, a newly formed group of NY-based companies, publishing houses, and researchers dedicated to advancing speech technology research and development.

show abstract

“…For an effective speech coder, the size of the EHMM must be considerably larger. Previous work [4][5][6] has shown that at least 64 states are required to represent all the acoustic variations present in fluent North American English. A single EHMM of this size can then be trained to model all of the sounds in such speech as well as the time dependent transitions from one sound to another.…”

Section: Ergodic Hidden Markov Modelsmentioning

confidence: 99%

Ultra Low Bit Rate Speech Coding Using an Ergodic Hidden Markov Model

Lee

Durey

Moore

et al.

Proceedings. (ICASSP '05). IEEE International Conference on Acoustics, Speech, and Signal Processing, 2005.

View full text Add to dashboard Cite

This paper presents the framework for an ultra low bit rate speech vocoder. The system is based on a recognition-synthesis paradigm in which a single ergodic hidden Markov model (EHMM) is used to capture the statistical characterizations of speech in a flexible manner capable of limiting the effects of recognition errors. Because predetermined speech units are not used, this system has the advantage of not requiring a transcription for the training data set. By incorporating a mixed excitation scheme based on an improved MELP formulation into the EHMM, additional gains in quality and speaker characterization are achieved at no cost to the bit rate.

show abstract

On the phonetic structure of a large hidden Markov model

Cited by 8 publications

References 4 publications

Phoneme-based vector quantization in a discrete HMM speech recognizer

Phoneme-based vector quantization in a discrete HMM speech recognizer

Ultra Low Bit-Rate Speech Coding

Ultra Low Bit Rate Speech Coding Using an Ergodic Hidden Markov Model

Contact Info

Product

Resources

About