nbest-mix

NAME

nbest-mix - interpolate N-best posterior probabilities

SYNOPSIS

nbest-mix [ -help ] option ... weight1 nbest1 weight2 nbest2 ...

DESCRIPTION

nbest-mix reads a number of N-best lists (which must contain identical hypotheses), computes the hypothesis posterior probabilities for each, and computes a new posterior distribution that is a weighted mixture of the input distributions. The hypothesis with the highest combined posterior probability is printed.

The command line arguments form an alternating list of weight values and N-best file names.

OPTIONS

Each filename argument can be an ASCII file, or a compressed file (name ending in .Z or .gz), or ``-'' to indicate stdin/stdout.

-help: Print option summary.
-version: Print version information.
-debug level: Controls the amount of output (the higher the level, the more).
-write-nbest file: Output N-best lists containing scores that correspond to the log of the combined posteriors of the input hypotheses. The log posterior is assigned as the acoustic score and other scores are set to zero. This also suppresses the printing of the best hyp.
-max-nbest n: Limits the number of hypotheses read from each N-best list to the first n.
-rescore-lmw lmw: Sets the language model weight used in combining the language model log probabilities with acoustic log probabilities (only relevant if separate scores are given in the N-best input).
-rescore-wtw wtw: Sets the word transition weight used to weight the number of words relative to the acoustic log probabilities (only relevant if separate scores are given in the N-best input).
-posterior-scale scale: Divide the total weighted log score by scale when computing normalized posterior probabilities. This controls the peakedness of the posterior distribution. The default value is whatever was chosen for lmw, so that language model scores are scaled to have weight 1, and acoustic scores have weight 1/lmw.
-set-lm-scores: In conjunction with -write-nbest, output N-best lists that preserve the acoustic scores and word counts of the (first of the) input N-best lists, and encodes the combined log posteriors via the LM scores. The LM scores in the output are calculated so that, when combined with the acoustic scores and insertion penalties (using the given LM weight and posterior scaling), the result is the weighted, combined posteriors based on all input N-best scores. This option is useful if input N-best lists were created by rescoring with different language models, and the output N-best lists are to be combined with other scores or if the score weighting is to be optimized with nbest-optimize(1).
-set-am-scores: Analogous to -set-lm-scores, except that the acoustic scores are modified to reflect combined log posterior probabiltities, and other scores are preserved. This option is useful if input N-best lists were created by rescoring with different acoustic models.

BUGS

Hopefully not.