Continuously variable duration hidden Markov models for automatic speech recognition

doi:10.1016/S0885-2308(86)80009-2

Computer Speech & Language

Volume 1, Issue 1, March 1986, Pages 29-45

https://doi.org/10.1016/S0885-2308(86)80009-2 Get rights and content

Abstract

During the past decade, the applicability of hidden Markov models (HMM) to various facets of speech analysis has been demonstrated in several different experiments. These investigations all rest on the assumption that speech is a quasi-stationary process whose stationary intervals can be identified with the occupancy of a single state of an appropriate HMM. In the traditional form of the HMM, the probability of duration of a state decreases exponentially with time. This behavior does not provide an adequate representation of the temporal structure of speech.

The solution proposed here is to replace the probability distributions of duration with continuous probability density functions to form a continuously variable duration hidden Markov model (CVDHMM). The gamma distribution is ideally suited to specification of the durational density since it is one-sided and only has two parameters which, together, define both mean and variance. The main result is a derivation and proof of convergence of re-estimation formulae for all the parameters of the CVDHMM. It is interesting to note that if the state durations are gamma-distributed, one of the formulae is non-algebraic but, fortuitously, has properties such that it is easily and rapidly solved numerically to any desired degree of accuracy. Other results are presented including the performance of the formulae on simulated data.

References (18)

BahlL.R. et al.
A maximum likelihood approach to continuous speech recognition
IEEE Transactions on Pattern Analysis and Machine Intelligence
(1983)
BakerJ.K.
The DRAGON.system—an overview
IEEE Transactions on Acoustics, Speech, and Signal Processing
(1975)
BaumL.E.
An inequality and associated maximization technique in statistical estimation for probabilistic functions of Markov processes
Inequalities
(1972)
BaumL.E. et al.
A maximization technique in the statistical analysis of probabilistic functions of Markov chains
Annals of Mathematical Statistics
(1970)
BaumL.E. et al.
Growth functions for transformations on manifolds
Pacific Journal of Mathematics
(1968)
BilliR.
Vector quantization and Markov source models applied to speech recognition
BourlardH. et al.
Connected digit recognition using vector quantization

There are more references available in the full text version of this article.

Cited by (362)

Sentiment analysis using novel and interpretable architectures of Hidden Markov Models[Formula presented]
2021, Knowledge-Based Systems
Citation Excerpt :
For many sequential observations, we anticipate that the trends in the data over several successive observations will provide important information in predicting the next value. The aforementioned restriction also does not allow combinations of states to be effectively modeled [53] and the self-loops on the states are poor duration models in most practical applications [54]. Quan et al. [43] found that utilizing more than one previous state increased performance and concluded that a sentence is clearly affected by as many as four previous sentences.
Sentiment analysis aims to formulate automated methods to recognize sentiments, opinions and emotions in text. Many methods and approaches have been utilized but most of them do not disclose the way that decisions are made and operate as black boxes. Hidden Markov Models (HMMs) constitute a quite suitable and potent approach for sentiment analysis, since they can utilize the sequential nature of the text, a piece of information that machine learning methods cannot properly utilize. However, little attention has been paid to formulating and applying sophisticated HMM-based methods and advanced training approaches for accomplishing sentiment analysis tasks. In this article, we introduce novel, interpretable HMM-based methods for recognizing sentiments in text and we examine their performance under various architectures, training methods, orders and ensembles. The introduced models possess interpretability, they can indicate the sentimental parts of a sentence and illustrate the way that the overall sentiment evolves from the start to the end of it. A concrete experimental study is conducted and the results show that the introduced HMMs methods and the training approaches are quite competitive with machine learning methods and that they outperform traditional HMMs. Furthermore, the designed HMMs methods possess great interpretability and can be an efficient approach for sentiment analysis.
Arabic audio clips: Identification and discrimination of authentic Cantillations from imitations
2020, Neurocomputing
This paper introduces a thorough study of classical and deep learning algorithms implemented for multi-class speaker identification and verification of Qur’anic audio clips. Thirty different reciters and twelve imitators, of top reciters, were evaluated in the study. In addition to identifying the reciter, our objective is to first evaluate different classifiers performing the stated recognition, compare classical vs. deep-based classifiers, and finally benchmark automatic against human expert listeners’ accuracy in identifying authentic reciters from imitators. Using different multimedia outlets over the internet, several reciters became more popular than others for their distinct cantillation style. Towards the development of a practical classifying system, a significant dataset of 15810 audio clips is constructed for thirty reciters in addition to 397 clips for top imitators. A combination of perceptual and acoustic features was extracted in order to achieve better classification. The classifying system is implemented using six top-achieving classical and two deep learning classifiers. Finally, we perform a survey engaging human expert listeners’ performance in detecting imitators from authentic reciters, against the fine-tuned classifiers to be able to discuss cross-comparative results. The concluded results demonstrated a high accuracy performance for the selected classifiers averaging $98.6 %$ on the testing dataset, $93 %$ on a separate testing dataset, $97 %$ on the problem of discriminating authentic Cantillations from imitations, and outstanding performance on the survey reaching an accuracy rate of $98 %$ compared to $61 %$ of human experts.
Latent state recognition by an enhanced hidden Markov model
2020, Expert Systems with Applications
In this paper, we start from relaxing assumptions of traditional hidden Markov model then develop a novel framework for decoding the latent states, from which the dynamics of multi-variable financial data is generated. To construct the framework, we model the observed variables as a p-order vector autoregressive process, allow the latent state to evolve through a semi-Markov chain, and shrink the auto-regression and covariance matrices via a penalized maximization likelihood method. Using the 50-dimensional simulated data, the 12-dimensional 5-min order book data of the Chinese CSI 300 index component stocks, the 49-dimensional daily data of U.S. industry portfolio, and 1-dimensional hourly data of four primary foreign exchange rates, our empirical analyses show that the proposed model outperforms the alternative model in accurately recognizing anomalous events and achieves better sharp ratio in a pseudo trading strategy via the latent states. The superior performance is across the data frequency of minute, hour and daily, the dimension of one, twelve, and fifty, the data type of stock, foreign exchange rate, and industry portfolio.
A hidden semi-Markov model for indoor radio source localization using received signal strength
2020, Signal Processing
Multipath propagation makes the use of received signal strength (RSS) unreliable as a signal propagation model for localization of a radio source based on RSS data. An approach to mitigation of this problem is the use of a Hidden Markov model (HMM) to represent the relationship between RSS and the radio source location by incorporating an environment prior and RSS source dynamics. The HMM structure forces a geometric form for the distribution for the sojourn time. This, combined with missing data problems, reduces confidence in location estimation. It is found, in this paper, that Hidden semi-Markov Models (HsMMs), with a more flexible sojourn time distribution are more able to represent source dynamics while retaining the advantages of HMMs for environmental constraints and improving resilience to missing measurements. The performance of the proposed HsMM algorithm is compared with a standard HMM via experiments and simulations involving indoor radio source localization using RSS data. Simulations use a ray tracing software-based simulator and the experiments for transmitter localization with RSS data are collected by software defined radios.
Rhythm Modeling for Voice Conversion
2023, arXiv
Inhomogeneous hidden semi-Markov models for incompletely observed point processes
2023, Annals of the Institute of Statistical Mathematics

View all citing articles on Scopus

View full text

Continuously variable duration hidden Markov models for automatic speech recognition

Abstract

A maximum likelihood approach to continuous speech recognition

IEEE Transactions on Pattern Analysis and Machine Intelligence

The DRAGON.system—an overview

IEEE Transactions on Acoustics, Speech, and Signal Processing

An inequality and associated maximization technique in statistical estimation for probabilistic functions of Markov processes

Inequalities

A maximization technique in the statistical analysis of probabilistic functions of Markov chains

Annals of Mathematical Statistics

Growth functions for transformations on manifolds

Pacific Journal of Mathematics

Vector quantization and Markov source models applied to speech recognition

Connected digit recognition using vector quantization