Hidden Markov Models: Difference between revisions

From MSc Voice Technology
Jump to navigation Jump to search
Line 3: Line 3:
== Introduction ==
== Introduction ==


A '''Hidden Markov Model (HMM)''' is a temporal probabilistic model in which some "hidden" or unobservable states are described by observable variables that are generated by these hidden states. <ref>Russell, S. J. (2010). ''Artificial intelligence a modern approach''. Pearson Education, Inc..</ref> These hidden states adhere to the Markov property, meaning that the future state is only dependent on the current state. Since one cannot observe the underlying states of a specific model, learning the transition function of this sequence of states involves aligning the HMM to the observable states.<ref>Eddy, S. R. (1996). Hidden markov models. ''Current opinion in structural biology'', ''6''(3), 361-365.</ref><ref name=":0">Rabiner, L. R. (1989). A tutorial on hidden Markov models and selected applications in speech recognition. ''Proceedings of the IEEE'', ''77''(2), 257-286.</ref>  
A '''Hidden Markov Model (HMM)''' is a temporal probabilistic model characterized by a sequence of hidden states, which are concealed from direct observation but influence observable variables generated by these states. <ref>Russell, S. J. (2010). ''Artificial intelligence a modern approach''. Pearson Education, Inc..</ref> These hidden states adhere to the Markov property, signifying that the future state depends solely on the current state. Given the unavailability of direct access to the underlying states in a specific model, the learning process involves aligning the HMM with observable states.<ref>Eddy, S. R. (1996). Hidden markov models. ''Current opinion in structural biology'', ''6''(3), 361-365.</ref><ref name=":0">Rabiner, L. R. (1989). A tutorial on hidden Markov models and selected applications in speech recognition. ''Proceedings of the IEEE'', ''77''(2), 257-286.</ref>  


Many real-world applications present hidden variables that are only observable through some emitted outcome, e.g. a speech signal of a word is observed rather than the specific phoneme states that are the underlying hidden states. To determine what the sequence of phonemes (states) would be that results in that specific word, the model learns the relation between the observed and unobservable variables.<ref>Juang, B. H., & Rabiner, L. R. (1991). Hidden Markov models for speech recognition. ''Technometrics'', ''33''(3), 251-272.</ref>  
In many real-world scenarios, concealed variables are only discernible through their emitted outcomes. For example, in speech recognition, we often observe the speech signal of a word rather than the specific phoneme states that constitute the hidden states. To decipher the sequence of phonemes leading to a particular word, HMMs facilitate the modeling of the relationship between observable and unobservable variables.<ref>Juang, B. H., & Rabiner, L. R. (1991). Hidden Markov models for speech recognition. ''Technometrics'', ''33''(3), 251-272.</ref>  


The technique behind Hidden Markov Models has been shown to be related to [[Dynamic Time Warping]].<ref>Juang, B. H. (1984). On the hidden Markov model and dynamic time warping for speech recognition—A unified view. ''AT&T Bell Laboratories Technical Journal'', ''63''(7), 1213-1243.</ref><ref>Fang, C. (2009). From dynamic time warping (DTW) to hidden markov model (HMM). ''University of Cincinnati'', ''3'', 19.</ref>  
Furthermore, the application of HMMs has exhibited connections with [[Dynamic Time Warping]], offering insights into their utility across diverse domains.<ref>Juang, B. H. (1984). On the hidden Markov model and dynamic time warping for speech recognition—A unified view. ''AT&T Bell Laboratories Technical Journal'', ''63''(7), 1213-1243.</ref><ref>Fang, C. (2009). From dynamic time warping (DTW) to hidden markov model (HMM). ''University of Cincinnati'', ''3'', 19.</ref>


== Historical Context ==
== Historical Context ==

Revision as of 09:07, 18 September 2023

Claimed by Ömer, Jocomin, and Ding.

Introduction

A Hidden Markov Model (HMM) is a temporal probabilistic model characterized by a sequence of hidden states, which are concealed from direct observation but influence observable variables generated by these states. [1] These hidden states adhere to the Markov property, signifying that the future state depends solely on the current state. Given the unavailability of direct access to the underlying states in a specific model, the learning process involves aligning the HMM with observable states.[2][3]

In many real-world scenarios, concealed variables are only discernible through their emitted outcomes. For example, in speech recognition, we often observe the speech signal of a word rather than the specific phoneme states that constitute the hidden states. To decipher the sequence of phonemes leading to a particular word, HMMs facilitate the modeling of the relationship between observable and unobservable variables.[4]

Furthermore, the application of HMMs has exhibited connections with Dynamic Time Warping, offering insights into their utility across diverse domains.[5][6]

Historical Context

The origin of Hidden Markov Models (HMM) dates back to 1907, where Andrei Markov formulated Markov chains, which had proved that dependent variables were also affected by the law of large numbers, rather than only independent variables, and was heavily influenced by the Bernoulli model.[7] Although this process was known for several decades, it was not until the 1960’s that Leonard Baum and Ted Petrie began to create a new model that would achieve the most likely estimate of the parameters of the Markov chain, further refining the probability equation to now find the hidden paths to the process.[8][9] Yet, it was Jelinek, Bahl, and Mercer who first utilized the Markov model in speech recognition, in their attempt to move away from speaker dependent probabilities, and has become one of the most common uses of HMM.[10][11]

Further amendments and improvements to combat frequent issues with the model have been made since, most notably in the 1980's and 1990's[12], including shared-distribution HMM, which more easily dealt with huge numbers of parameters with limited training data[13], Hierarchical Hidden Markov Models, which generalized standard HMMs and made the hidden states autonomous models, leading to sequences rather than single symbols being output.[14], and signal decomposition, where parallel HMMs are used to simultaneously recognise concurrent events, e.g. separating background noise from speech.[15]

In later years, namely the latter half of the 1980’s, HMM was being used for DNA sequencing and biological computations.[16]

Key Innovations

Some key innovations in the field of speech recognition using Hidden Markov Models include [17]:

  • The DRAGON System developed by Dr. James Baker was one of the earlier speech recognition systems that used HMMs and later became known as Dragon Dictate.[18] DRAGON is a probabilistic model that represents all knowledge from the training set of utterances in a transition matrix and a matrix of conditional probabilities between the hidden states and the observable states. This allowed the system to be speaker-agnostic, while being quick in finding the optimal path of recognition through dynamic programming. Carnegie Mellon's Harpy System improved upon DRAGON by incorporating speech-dependent heuristics and other improvements to increase the performance. [19]
  • DARPA Speech Understanding Research funded multiple laboratories in speech recognition, including BYBLOS and SPHINX, both using HMMs. [20][21]
  • Many voice assistants used HMMs before the Deep Learning Revolution and the development of end-to-end models, an example is Siri. [22]

All in all, the impact of HMMs in speech recognition has been significant as faster, more simplified, and more generalized alternatives to conventional knowledge representation models.

Impact on the Field

Hidden Markov Models (HMMs) have left an indelible mark on the field of speech recognition, shaping its evolution and expanding its horizons. Their impact is felt across various dimensions of speech processing, underscoring their enduring relevance and adaptability.

  • HMMs have been a cornerstone of large vocabulary speech recognition, revolutionizing the way extensive vocabularies are processed. This inherent capability has been instrumental in advancing the accuracy and efficiency of large vocabulary speech recognition systems, contributing to their widespread adoption.[3]
  • In 1996, Bourlard and Dupont marked a significant milestone by employing HMMs in the domain of automatic subtitle generation. Their pioneering work showcased the versatility of HMMs in transcribing spoken content into textual subtitles. This breakthrough enhanced accessibility and facilitated the dissemination of audiovisual content, making it more inclusive and accessible to diverse audiences.[23]
  • In the domain of continuous speech recognition , these models demonstrate notable expertise in the representation of acoustic features inherent to speech and the formulation of language models conducive to the process of decoding spoken utterances. The incorporation of HMMs within continuous speech recognition systems has played a pivotal role in facilitating the attainment of robust and precise transcription capabilities. Consequently, this integration has substantially broadened the scope of applications within the field of voice recognition and transcription.[24]
  • While the landscape of speech recognition has evolved with the advent of deep learning and neural networks, HMMs have not faded into obsolescence. In 2021, MIT researchers showcased the enduring importance of HMMs by integrating them into a speaker-adapted HMM-GMM (Hidden Markov Model-Gaussian Mixture Model) Automatic Speech Recognition (ASR) system. This groundbreaking research involved training HMMs using audio data and source language VTT subtitles, leveraging the KALDI framework and Mel-Frequency Cepstral Coefficients (MFCC) features. The outcome was a model that seamlessly combined speech recognition and syntax translation, highlighting the adaptability and effectiveness of HMMs in addressing complex tasks at the intersection of speech processing and natural language understanding.[25] (Honestly, I don't fully understand all the technicalities of the "25th literature", but HMM played an important role in their research.)

Future Research

LLM Review

References

Here thus are the references: [26]

  1. Russell, S. J. (2010). Artificial intelligence a modern approach. Pearson Education, Inc..
  2. Eddy, S. R. (1996). Hidden markov models. Current opinion in structural biology, 6(3), 361-365.
  3. 3.0 3.1 Rabiner, L. R. (1989). A tutorial on hidden Markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2), 257-286.
  4. Juang, B. H., & Rabiner, L. R. (1991). Hidden Markov models for speech recognition. Technometrics, 33(3), 251-272.
  5. Juang, B. H. (1984). On the hidden Markov model and dynamic time warping for speech recognition—A unified view. AT&T Bell Laboratories Technical Journal, 63(7), 1213-1243.
  6. Fang, C. (2009). From dynamic time warping (DTW) to hidden markov model (HMM). University of Cincinnati, 3, 19.
  7. Gagniuc, P.A. (2017). Historical Notes. In Markov Chains, P.A. Gagniuc (Ed.). https://doi.org/10.1002/9781119387596.ch1
  8. Baum, Leonard E., and Ted Petrie. “Statistical Inference for Probabilistic Functions of Finite State Markov Chains.” The Annals of Mathematical Statistics 37, no. 6 (December 1966): 1554–63. https://doi.org/10.1214/aoms/1177699147.
  9. Nilsson, Mikael, and Marcus Ejnarsson. “Speech Recognition Using Hidden Markov Model,” n.d.
  10. Jelinek, F., L. Bahl, and R. Mercer. “Design of a Linguistic Statistical Decoder for the Recognition of Continuous Speech.” IEEE Transactions on Information Theory 21, no. 3 (May 1975): 250–56. https://doi.org/10.1109/TIT.1975.1055384.
  11. Stamp, Mark. “A Revealing Introduction to Hidden Markov Models.” In Introduction to Machine Learning with Applications in Information Security, by Mark Stamp, 7–35, 1st ed. Chapman and Hall/CRC, 2017. https://doi.org/10.1201/9781315213262-2.
  12. Gales, Mark, and Steve Young. “The Application of Hidden Markov Models in Speech Recognition.” Foundations and Trends® in Signal Processing 1, no. 3 (February 20, 2008): 195–304. https://doi.org/10.1561/2000000004.
  13. Hwang, Mei-Yuh, and Xuedong Huang. “Shared-Distribution Hidden Markov Models for Speech Recognition.” IEEE Transactions on Speech and Audio Processing 1, no. 4 (October 1993): 414–20. https://doi.org/10.1109/89.242487.
  14. Fine, Shai, Yoram Singer, and Naftali Tishby. “The Hierarchical Hidden Markov Model: Analysis and Applications.” Machine Learning 32, no. 1 (July 1, 1998): 41–62. https://doi.org/10.1023/A:1007469218079.
  15. Varga, A.P., and R.K. Moore. “Hidden Markov Model Decomposition of Speech and Noise.” In International Conference on Acoustics, Speech, and Signal Processing, 845–48. Albuquerque, NM, USA: IEEE, 1990. https://doi.org/10.1109/ICASSP.1990.115970.
  16. Eddy, Sean R. “What Is a Hidden Markov Model?” Nature Biotechnology 22, no. 10 (October 2004): 1315–16. https://doi.org/10.1038/nbt1004-1315.
  17. Juang, B. H., & Rabiner, L. R. (1991). Hidden Markov models for speech recognition. Technometrics, 33(3), 251-272.
  18. Baker, J. (1975). The DRAGON system--An overview. IEEE Transactions on Acoustics, speech, and signal Processing, 23(1), 24-29.
  19. Lowerre, B. T. (1976). The Harpy speech recognition system [Ph. D. Thesis].
  20. Chow, Y., Dunham, M., Kimball, O., Krasner, M., Kubala, G., Makhoul, J., ... & Schwartz, R. (1987, April). BYBLOS: The BBN continuous speech recognition system. In ICASSP'87. IEEE International Conference on Acoustics, Speech, and Signal Processing (Vol. 12, pp. 89-92). IEEE.
  21. Lee, K. F. (1988). Automatic speech recognition: the development of the SPHINX system (Vol. 62). Springer Science & Business Media.
  22. Domingos, P. (2015). The master algorithm: How the quest for the ultimate learning machine will remake our world. Basic Books.
  23. H. Bourlard and S. Dupont, “A new ASR approach based on independent processing and recombination of partial frequency bands.” Proc. of the IEEE International Conference on Acoustics, Speech, and Signal Processing, Vol. 1, pages 426--429, Philadephia, Pennsylvania, October 1996.
  24. Young S J, Evermann G, Gales M J F, et al. The HTK book version 3.4 Manual[J]. Cambridge University Engineering Department, Cambridge, UK, 2006.
  25. Salesky E, Wiesner M, Bremerman J, et al. The multilingual tedx corpus for speech recognition and translation[J]. arXiv preprint arXiv:2102.01757, 2021.
  26. Placeholder Reference