Joseph A. Giampapa
6 November 2014 garof@cs.cmu.edu www.cs.cmu.edu/~garof
Background of HMMs
• Best early tutorial:
Lawrence R. Rabiner, “A Tutorial on Hidden Markov Models and
Selected Applications in Speech Recognition,” in Proceedings of the
IEEE, Vol. 77, No. 2, February 1989. DOI: 0018-9219/89/020002,
URL:
http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=18626
• HMMs
– Introduced and studied in late 1960s and 1970s
– Became popular in the late 1980s
• Reasons for their popularity
– “Rich in mathematical structure” (Rabiner)
– They work well in practice for certain types of applications
– In particular, useful for characterizing signal models
Signal Model
• Consists of two parts:
– The observations, in a time-dependent sequence
• This is the signal
• Observation sequence can be discrete or continuous
– Discrete: alphabet, samples, sample intervals
– Continuous: speech, temperature, power
– The process that produces the observations
• Reasons for having a signal model:
– To understand and/or simulate the process
– To identify and recognize the signal
– To filter, transform, break-apart, componentize
(process) signals
Types of Signal Models
• Deterministic model
– E.g. sine wave, sum of exponentials
– The equation that describes it is known
– Just supply the parameters
• E.g. amplitude, frequency, phase
• Statistical model
– The sequence of observables is not easily characterized by a deterministic description
– You can hypothesize the variables and known quantities that influence the signal’s generation
– Underlying assumption: whatever produces the signal can be described by a parametric random process
• Question: How would you characterize the power consumption of
the