HMMDecoderMD

The HMMDecoderMD is an extended version of the previously described HMMDecoderStandard that has been adapted to handle missing data. As well as an input stream of feature vectors, it also accepts a `missing data mask' (in2) and optional lower and upper bound values (in3 and in4). The vectors composing the mask and the bounds should have the same width as the feature vectors. The missing data mask should be a vector of boolean values. Each boolean element dictates whether the corresponding feature vector element is to be considered to be present (i.e. to contain reliable data) or to be missing (i.e. to contain unreliable data). Present values are indicated by 1's, and missing values by 0's.

The HMMDecoderMD has the same parameters as the HMMDecoderStandard, plus the two additional parameters that are described below:

• USE_BOUNDS

  This is a boolean parameter that determines whether or not the decoder should appled bounded marginalisation to the missing components of the input data. When set to false the employs the marginal distribution for the present data constructed by `integrating out' the missing values allowing them to lie anywhere between + and - infinity. This is extremely quick but produces poor results when there is a lot of data missing. When the USE_BOUNDS switch is set to true the decoder will use the so-called `bounds constraint' for the missing components of the input data. This means that instead of treating the missing data as completely unknown, the missing data is taken to be constrained between a lower and an upper bound value. For spectral energies the lower bound is typically 0 and the upper bound is typically the observed energy in the noisy signal i.e. logically, if data is flagged as missing, then the energy of the clean signal at that particular spectro-temporal point must be positive but must also be sufficiently low to have been masked by the observed energy at that point. In terms of computation, rather than integrating the distributions for the missing components between -infinity and +infinity, they are now integrated between the lower and upper bound. The integration is performed using the Gaussian error function. Although evaluating this integration imposes a heavier computational load, it can lead to greatly improved recognition performance.

  The upper and lower bounds values are taken from the optional inputs in3 and in4 respectively. If the bound inputs are not connected then the lower bounds value defaults to 0 and the upper bound defaults to the input feature value (i.e. in1).

  Note, the USE_BOUNDS switch causes bounded marginalisation to be applied to the static (i.e. non-delta) features. To extend the use of bounded marginalisation to the delta features you must also set the USE_DELTA_BOUNDS switch (see below).

• USE_DELTA_BOUNDS

  This switch is similar to the DELTA_BOUDS switch, but it controls whether or not bounded marginalisation is applied to the delta features. The bounded marginalisation can only be applied to delta features if it is also being applied to the non-delta features (i.e. USE_DELTA_BOUNDS can only be set to true if USE_BOUNDS is also true.)

  If bounded marginalisation is applied to delta features then it is necessary to supply appropriate bound using inputs in3 and in4. Suitable bounds for delta features can be computed using the DeltaBounds block.

As with HMMDecoderStandard, the missing data decoder outputs a stream of state-likelihood frames and a stream of mixture labels. Each state-likelihood frame consists of the likelihood of each model state having generated the corresponding input feature frame. Within these frames the state likelihoods occur in the same order in which the states are defined in the HMM definition file. (The mixture label frames (out2) indicate the integer label of the winning mixture for each state).

Inputs	Meaning	Sample	1-D frame	2-D frame
in1	feature vectors	No	Yes	No
in2	missing data mask	No	Yes	No
(in3)	lower bound	No	Yes	No
(in4)	upper bound	No	Yes	No

Outputs	Meaning
out1	state likelihoods
out2	state max mixture label

Parameters	Type	Default	Meaning
LOG_FILE	String	-	Name of an optional log file
LOG_FILE_2	String	-	Name of additional detailed log file
WORD_PENALTY	Float	0.0	The creation penalty
HMM_FILE	String	-	Name of the HMM file list
GRAMMAR_FILE	String	-	File storing the grammar
LABEL_FILE	String	-	File storing HMM NAME-> HMM LABEL mapping
FIRST_TOKEN	String	-	Label of a fixed first token
FINAL_TOKEN	String	-	Label of a fixed final token
TRANSCRIPTION	String	-	The correct transcription
SILENCE	String	""	The silence label(s)
MAX_APPROX	Boolean	False	Use max mixture approximation
NBEST	Int	1	Return best N hypotheses
STATE_PATH	Boolean	False	Record HMM state path
HAS_DELTAS	Boolean	0	Models have delta parameters
USE_DELTAS	Boolean	-	Models have delta parameters
HYPOTHESIS FILTER	String	""	Regular expression for filtering hypotheses
OUTPUT_CONFUSIONS	Boolean	0	Output confusion matrix
DUMP_PARAMETERS	Boolean	0	Write parameters to log file
USE_BOUNDS	Boolean	False	Use bounded marginalisation (static features)
USE_DELTA_BOUNDS	Boolean	False	Use bounded marginalisation (delta features)