HMMDecoderMultisource

The `multisource' decoder accepts a 1/0 missing data mask and a mask of integer labels that define grouped regions (`fragments') in the 1/0 mask. These fragments are assumed to be regions of the representation that are due to a single source. Rather than using a single fixed present data mask the decoder attempts to test every mask that can be generated from a subset of the mask fragments. As there is potentially a very large number of mask hypotheses, a limited search is employed. This search proceeds by generating new complimentary hypotheses when a fragment starts and merging partial complementary hypotheses every time a fragment ends. The algorithm is described in more details in:

• J. Barker, M.P. Cooke and D.P.W.Ellis (2000). Decoding speech in the presence of other sound sources, Proc. ICSLP-00, Beijing, China, PDF | PostScript

The multisource decoder has five inputs. The data, mask, group labellings and lower and upper missing data bounds.

The parameters are similar to those of HMMDecoderMD but with the following additions:

• DISPLAY_GROUPS

  A boolean parameter that if set to TRUE causes the mask that was used for the winning hypothesis to be displayed.

• MD_WEIGHT

  A parameter that controls the ratio of present/missing data in the mask. The exact interpretation of this parameter depends on the normalisation technique the decoder employs. The normalisation technique is set using the NORMALISE_MODE parameter.

• NORMALISE_MODE

  The current implementation can employ two different techniques for balancing the amount of present against the amount of missing data. The user can selected between these by setting the NORMALISE_MODE parameter to either MODE_1 or MODE_2:

  • MODE_1

    When operating in MODE_1 all missing data probabilities are scaled by the parameter MD_WEIGHT. Present data probabilities are not altered. So, with a higher value of MD_WEIGHT the decoding will be biased towards mask hypotheses in which more data is missing, and with a lower value it will favour hypotheses in which less data is missing.

  • MODE_2

    In this mode, individual present and missing data probabilities are employed unscaled, but entire mask hypothesis probabilities are adjusted at the point where a fragment ends and complimentary hypotheses are compared. The hypothesis score adjustment is based on the size of the fragment that is under consideration. The score for the mask hypothesis in which the fragment is considered to be present remains unaltered, and the score for the hypothesis in which the fragment has been dropped is reduced. The reduction in the log probability score is computed as the size of the fragment multiplied by the MD_WEIGHT parameter. Hence, a large value MD_WEIGHT implies a larger cost for labelling fragments as missing.

  MD_WEIGHT has very different interpretations in modes 1 and 2. Note, a larger MD_WEIGHT biases towards more missing data in MODE_1 but towards less missing data in MODE_2.

Inputs	Meaning	Sample	1-D frame	2-D frame
in1	feature vectors		Yes
in2	fuzzy data mask		Yes
in3	group labels		Yes
(in4)	lower bound		Yes
in5	upper bound		Yes

Outputs	Meaning
out1	rubbish!
out2	rubbish!

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
DISPLAY_GROUPS	Boolean	0	Display mask backtrace (requires MATLAB)
USE_DELTA_BOUNDS	Boolean	False	Use bounded marginalisation (delta features)
MD_WEIGHT	Float	0.0	Balances amount of present/missing data
NORMALISE_MODE	{MODE_1, MODE_2}	MODE_1	The technique used to balance the amount of present/missing data