Explicit Duration Research Articles

Acoustic modeling problem for automatic speech recognition system: advances and refinements (Part II)

In automatic speech recognition (ASR) systems, hidden Markov models (HMMs) have been widely used for modeling the temporal speech signal. As discussed in Part I, the conventional acoustic models used for ASR have many drawbacks like weak duration modeling and poor discrimination. This paper (Part II) presents a review on the techniques which have been proposed in literature for the refinements of standard HMM methods to cope with their limitations. Current advancements related to this topic are also outlined. The approaches emphasized in this part of review are connectionist approach, explicit duration modeling, discriminative training and margin based estimation methods. Further, various challenges and performance issues such as environmental variability, tied mixture modeling, and handling of distant speech signals are analyzed along with the directions for future research.

Transport of Selected PET Radiotracers by Human P-Glycoprotein (ABCB1) and Breast Cancer Resistance Protein (ABCG2): An In Vitro Screening

Radiolabeled compounds used for brain imaging with PET must readily cross the blood-brain barrier (BBB) to reach their target. Efflux transporters at the BBB-P-glycoprotein (P-gp) and the breast cancer resistance protein (BCRP)-could limit their uptake by the brain. We developed and validated an in vitro model using MDCKII cells transfected with human multidrug resistance (MDR1) or BCRP genes and assessed the transport of selected PET ligands by the concentration equilibrium technique. The tested compounds included befloxatone, (R,S)-CGP-12177, clorgyline, R-(-)-deprenyl, diprenorphine, DPA-714, fallypride, flumazenil, 2-fluoro-A-85380, LBT-999, loperamide, p-MPPF, PE2I, Pittsburgh compound B (PIB), (R,S)-PK11195, raclopride, R-(+)-verapamil, and WAY-100635. The assays were performed using the nonradioactive form of each compound (ultraviolet high-performance liquid chromatography analysis) and, when available, the (18)F-labeled analogs (γ-counting). Befloxatone appeared to be transported solely by BCRP. Loperamide, verapamil, and diprenorphine were the only P-gp substrates. Other ligands were transported by neither P-gp nor BCRP. The present method can readily be used to screen new-compound transport by P-gp or BCRP, even before any radiolabeling. Compounds that were previously thought to be transported by P-gp in rodents, such as p-MPPF, WAY-100635, and flumazenil, cannot be considered substrates of human P-gp. The impact of BCRP and P-gp at the BBB on the transport of befloxatone and diprenorphine in vivo remains to be evaluated with PET.

Average-Voice-Based Speech Synthesis Using HSMM-Based Speaker Adaptation and Adaptive Training

In speaker adaptation for speech synthesis, it is desirable to convert both voice characteristics and prosodic features such as F0 and phone duration. For simultaneous adaptation of spectrum, F0 and phone duration within the HMM framework, we need to transform not only the state output distributions corresponding to spectrum and F0 but also the duration distributions corresponding to phone duration. However, it is not straightforward to adapt the state duration because the original HMM does not have explicit duration distributions. Therefore, we utilize the framework of the hidden semi-Markov model (HSMM), which is an HMM having explicit state duration distributions, and we apply an HSMM-based model adaptation algorithm to simultaneously transform both the state output and state duration distributions. Furthermore, we propose an HSMM-based adaptive training algorithm to simultaneously normalize the state output and state duration distributions of the average voice model. We incorporate these techniques into our HSMM-based speech synthesis system, and show their effectiveness from the results of subjective and objective evaluation tests.

Prosody dependent speech recognition on radio news corpus of American English

Does prosody help word recognition? This paper proposes a novel probabilistic framework in which word and phoneme are dependent on prosody in a way that reduces word error rates (WER) relative to a prosody-independent recognizer with comparable parameter count. In the proposed prosody-dependent speech recognizer, word and phoneme models are conditioned on two important prosodic variables: the intonational phrase boundary and the pitch accent. An information-theoretic analysis is provided to show that prosody dependent acoustic and language modeling can increase the mutual information between the true word hypothesis and the acoustic observation by exciting the interaction between prosody dependent acoustic model and prosody dependent language model. Empirically, results indicate that the influence of these prosodic variables on allophonic models are mainly restricted to a small subset of distributions: the duration PDFs (modeled using an explicit duration hidden Markov model or EDHMM) and the acoustic-prosodic observation PDFs (normalized pitch frequency). Influence of prosody on cepstral features is limited to a subset of phonemes: for example, vowels may be influenced by both accent and phrase position, but phrase-initial and phrase-final consonants are independent of accent. Leveraging these results, effective prosody dependent allophonic models are built with minimal increase in parameter count. These prosody dependent speech recognizers are able to reduce word error rates by up to 11% relative to prosody independent recognizers with comparable parameter count, in experiments based on the prosodically-transcribed Boston Radio News corpus.

Discriminative training and explicit duration modeling for HMM-based automatic segmentation

HMM-based automatic segmentation has been popularly used for corpus construction for concatenative speech synthesis. Since the most important reasons for the inaccuracy of HMM-based automatic segmentation are the HMM training criterion and duration control, we will study these particular issues. For the HMM training, we apply the discriminative training method and introduce a new criterion, named Minimum SeGmentation Error (MSGE). In this method, a loss function directly related to the segmentation error is defined, and parameter optimization is performed by the Generalized Probabilistic Descent (GPD) algorithm. For the duration control problem, we apply explicit duration models and propose a two-step-based segmentation method to solve the problem of computational cost, where the duration model is incorporated in a postprocessor procedure. From the experimental results, these two techniques significantly improve segmentation accuracy with different focuses, where the MSGE-based discriminative training focuses on improving the accuracy of sensitive boundary, i.e., a boundary where an error in segmentation is likely to cause a noticeable degradation in speech synthesis quality, and the explicit duration modeling focuses on eliminating large errors. After combining these two techniques, the error average was reduced from 6.86 ms to 5.79 ms on Japanese data, and from 8.67 ms to 6.61 ms on Chinese data. Simultaneously, the number of errors larger than 30 ms were reduced 25% and 51% on Chinese and Japanese data, respectively.

An efficient forward-backward algorithm for an explicit-duration hidden Markov model

Existing algorithms for estimating the model parameters of an explicit-duration hidden Markov model (HMM) usually require computations as large as O((MD/sup 2/ + M/sup 2/)T) or O(M/sup 2/ DT), where M is the number of states; D is the maximum possible interval between state transitions; and T is the period of observations used to estimate the model parameters. Because of such computational requirements, these algorithms are not practical when we wish to construct an HMM model with large state space and large explicit state duration and process a large amount of measurement data to obtain high accuracy. We propose a new forward-backward algorithm whose computational complexity is only O((MD + M/sup 2/)T), a reduction by almost a factor of D when D > M and whose memory requirement is O(MT). As an application example, we discuss an HMM characterization of access traffic observed at a large-scale Web site: we formulate the Web access pattern in terms of an HMM with explicit duration and estimate the model parameters using our algorithm.

On the complexity of explicit duration HMM's

Introduces a new recursion that reduces the complexity of training a semi-Markov model with continuous output distributions. It is shown that the cost of training is proportional to M/sup 2/+D, compared to M/sup 2/D with the standard recursion, where M is the observation vector length and D is the maximum allowed duration. >

A Parallel Implementation of a Hidden Markov Model with Duration Modeling for Speech Recognition

Hidden Markov models (HMMs) are currently the most successful paradigm for speech recognition. Although explicit duration continuous HMMs more accurately model speech than HMMs with implicit duration modeling, the cost of accurate duration modeling is often considered prohibitive. This paper describes a parallel implementation of an HMM with explicit duration modeling for spoken language recognition on the MasPar MP-1. The MP-1 is a fine-grained SIMD architecture with 16384 processing elements (PEs) arranged in a 128 × 128 mesh. By exploiting the massive parallelism of explicit duration HMMs, development and testing is practical even for large amounts of data. The result of this work is a parallel speech recognizer that can train a phone recognizer in real time. We present several extensions that include context dependent modeling, word recognition, and implicit duration HMMs.

VOICE: An integrated speech recognition synthesis system for the Hindi language

A Voice Oriented Interactive Computing Environment (VOICE) has been implemented in the Hindi language. The system provides in interactive facility for visual and voice feedback. The 200 isolated word recognition system is designed around a railway reservation enquiry task and uses acoustic-phonetic segments as the basic units of recognition. Frame level classification into broad acoustic-phonetic categories is accomplished by a maximum likelihood classifier and segmentation by hierarchical clustering of the frame level likelihood vectors by use of explicit duration semi (Hidden) Markov Models. A more detailed classification of a few categories (vowels, voice bar and nasals in the first instance) is performed by neural nets. String matching using dynamic programming accomplishes lexical access, or conversion of the phonetic category symbol strings into words. Distributed processing of the word recognition task enables recognition at four times real time. A language processor disambiguates between multiple choices given by the recognizer for each word and even corrects some acoustic level recognition errors. This, the first system working in any Indian language, gives a recognition performance of 85% at the word level. For comparison, a purely HMM based word level recognizer has also been implemented. The performance is expected to improve further as there is still substantial scope for refinement.

Use of explicit duration models in speech recognition

Although researchers have observed that automatic speech recognition errors are often associated with improbable duration patterns, current recognition systems take little advantage of known durational cues. In addition to using simplistic models of duration, recognizers put very little weight on durational probabilities (because of the high dimensionality of the feature vectors typically used). This work explores the benefits that can be gained in speaker‐independent continuous speech recognition from more accurate duration modeling and automatically estimating a duration score weight. In the context of the stochastic segment model, segment durations are modeled using gamma distributions conditioned on phonetic, lexical, and phrasal context, as well as speaking rate. Robust parameter estimation, a potential problem with the large number of different conditioning factors, is achieved using automatic clustering techniques with a maximum likelihood criterion. A by‐product of the clustering procedure is a better understanding of the relative importance of the conditioning factors. Speech recognition experiments on the resource management task show that simple context‐independent gamma distributions give similar results to nonparametric relative frequency models, and that error can be reduced by more than 10% by conditioning models on such factors as prepausal position, speaking rate, and lexical stress.