The BBN BYBLOS continuous speech recognition system uses 
context-dependent phonetic hidden Markov models (HMM) to achieve
state-of-the-art recognition performance on large vocabulary
applications [1].


The system used to generate results for the designated Oct 89
evaluation test had the following notable features:

(1) Speech input to the system was represented by 3 independent
    8-bit codebooks made from the following features:
    - 14 Mel scale warped cepstra (C1--C14), at 10 ms frame rate.
    - 14 delta-cepstra estimated by a linear regression on
      each cepstral feature over a 50 ms window.
    - Normalized energy and delta-energy (2 features total).
    For decoding, each codebook was represented as a mixture of
    Gaussians which was shared by all states in the HMMs.

(2) Distinct models were trained for the phonemes in the dictionary
    as well as for all the left diphones, right diphones,
    and triphones (including those occurring across word boundaries)
    found in the training data.
    The parameters of the context-dependent models were interpolated
    together with the context-independent phoneme models.
    Interpolation weights (lambdas) were computed as a function of 
    context type, state within a model, and number of occurrences 
    of the context in training [2].

(3) A 'triphone co-occurrence' smoothing was applied to the
    context-dependent observation densities after training [3].


References:

[1] Schwartz, R., C. Barry, Y. Chow, A. Derr, M. Feng, O. Kimball,
    F. Kubala, J. Makhoul, J. Vandegrift, "The BBN BYBLOS Continuous
    Speech Recognition System", Proceedings of the DARPA Speech and
    Natural Language Workshop, Philadelphia, Pennsylvania, 
    February 1989, pp. 94-99.
    
[2] Schwartz, R., Y. Chow, O. Kimball, S. Roucos, M. Krasner, 
    J. Makhoul, "Context-Dependent Modeling for Acoustic-Phonetic
    Recognition of Continuous Speech", IEEE Int. Conf. Acoustics, 
    Speech, Signal Processing, Tampa Florida, March 1985, pp. 1205-1208.

[3] Schwartz, R., O. Kimball, F. Kubala, M. Feng, Y. Chow, C. Barry,
    J. Makhoul, "Robust Smoothing Methods for Discrete Hidden Markov 
    Models", IEEE Int. Conf. Acoustics, Speech, Signal Processing,
    Glasgow, Scotland, May 1989, pp. 548-551.