Phoneme HMMs constrained by frame correlations

Takahashi, Satoshi; Matsuoka, Tatsuo; Minami, Yasuhiro; Shikano, Kiyohiro

doi:10.1109/icassp.1993.319274

Cited by 24 publications

(7 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our IFDHMM is compared with the standard HMM and with the linear-predictive HMM (LPHMM) (Kenny et al, 1990), which is equivalent to a continuous version of the bigram HMM (Wellekens, 1987;Paliwal, 1993;Takahashi et al, 1993) when the prediction order is set to one. The LPHMM can be viewed as a typical example of those models in which a joint Gaussian distribution representation is assumed in the modelling of the segments of speech frames.…”

Section: Methodsmentioning

confidence: 99%

“…In Equation (3) we assume that the probability of occurrence of s is independent of . The bigram HMM (Wellekens, 1987;Paliwal, 1993;Takahashi et al, 1993) can be included as a special case of Equation (3) by setting the order of dependence, N, to be unity and by defining the corresponding time-lag of the single conditional frame to be some constant. Previous less general efforts of optimizing the time-lags for such a model, using the same maximum-likelihood principle, have been reported by the authors (Ming & Smith, 1994;Smith et al, 1995).…”

Section: The Modelmentioning

confidence: 99%

“…Looking for alternatives, extensive studies have been carried out towards the modelling of the spectral dynamics of speech in an HMM framework. These studies can be traced, typically, from the bigram or N-gram constrained HMM (Wellekens, 1987;Paliwal, 1993;Takahashi, Matsuoka, Minami & Shikano, 1993;Nakagawa & Yamamoto, 1996), to the linear-predictive HMM (Kenny, Lennig & Mermelstein, 1990;Woodland, 1992), and to a large family of segment-based HMMs (see, for example, Brown, 1987;Russell, 1993;Bahl, de Souza, Gopalakrishnan, Nahamoo & Picheny, 1994;Milner & Vaseghi, 1994;Zavaliagkos, Zhao, Schwartz & Makhoul, 1994;Nakagawa & Yamamoto, 1996). In the bigram, N-gram or linear-predictive HMMs, each observed frame is made to be explicitly dependent on one or more previous frames by using some type of conditional observation densities; and in the segment-based HMMs, segments of frames are handled by using a neural net (Zavaliagkos et al, 1994), by assuming some type of conditional independence (Russell, 1993), or by performing some transformations to reduce the data size and/or to achieve some robustness of the resulting parameters (Bahl et al, 1994;Milner & Vaseghi, 1994;Nakagawa & Yamamoto, 1996).…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Modelling of the interframe dependence in an HMM using conditional Gaussian mixtures

Smith

1996

Computer Speech & Language

View full text Add to dashboard Cite

Section: Methodsmentioning

confidence: 99%

Section: The Modelmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Modelling of the interframe dependence in an HMM using conditional Gaussian mixtures

Smith

1996

Computer Speech & Language

View full text Add to dashboard Cite

“…A speech signal representing a symbol can be represented, in a parametric space, as a point which moves as articulatory configuration changes [19,20,21]. The trace of this moving point is called the trajectory of the symbol.…”

Section: 1mentioning

confidence: 99%

A statistical model for an automatic procedure to compress a word transcription dictionary

Mouria-Beji¹

1998

Advances in Pattern Recognition

View full text Add to dashboard Cite

Abstractvarious experiments have conclusively shown that superior continuous speech recognition performance is obtained when using context-dependent phonemic models. However, we have observed that using an explicit context-dependent phonemic model can yield many transcriptions for a single lexicon entry. In this work, we study the compression of the word transcription dictionaries (WTD) into a more compact form to balance the need between flexibility and reliability. Based on a measure of a likelihood function, a statistical model for an automatic procedure to compress a WTD is developed. The compressed dictionary is then used for sentence recognition in a continuous speech recognition system. Experimental results indicate a substantial improvement of the recognition rate after compression. IntroductionSeveral continuous speech recognition systems currently being developed use context-dependent models which seek to capture the pronunciation variations resulting from phonetic context effects. It has been observed that with these models, large vocabulary speech recognition systems usually necessitates to compress, from some standard references, a word phonetic transcription dictionary (WTD) [i, 2, 3, 4]. Such a dictionary generally gives a single transcription for a lexicon entry. Continuous speech recognition systems have shown satisfactory results when using dictionaries prepared in this way [3,5,6,7]. However, because of the large variations in the pronouneiation of a given word [8, 9, 10, 11, 12, 13, ], it is very difficult, when compressing a dictionary, to capture its most representative variant. Word transcriptions are first obtained from the context-dependent phonemic model CODEPHON-STM based on tile automatically expending speed and context (AESC) approach developed in our laboratory [11,4], and then compressed into a more compact form to balance the need between flexibility and reliability using an automatic procedure. In fact, the direct word transcriptions given by CODEPHON-STM can yield many transcriptions for a single lexicon entry.

show abstract

“…Then, an algorithm to find the best state sequence in the HSMM was defined, aiming for a more explicit modeling of context. In another approach to overcome the temporal limitations of the standard HMM framework, alternative trajectory modeling (Takahashi, 1993) has been proposed, taking advantage of frame correlation. The models obtained can improve speech recognition performance, but they generally require a demoralizing increase in model parameters and computational complexity.…”

Section: Time Independence and Parameter Independence Assumptionsmentioning

confidence: 99%