nbest-format - File formats for N-best hypotheses lists


SRILM currently understands three different formats for lists of N-best hypotheses for rescoring or 1-best hypothesis extraction. The first two formats originated in the SRI Decipher(TM) recognition system, the third format is particular to SRILM.

The first format consists of the header

followed by one or more lines of the form
	(score) w1 w2 w3 ...
where score is a composite acoustic/language model score from the recognizer, on the bytelog scale. (A bytelog is a logarithm to base 1.0001, divided by 1024 and rounded to an integer.) This format is output by the SRI Decipher(TM) recognizer, by the ngram -nbest, and by nbest-lattice -write-nbest -decipher-nbest.

The second Decipher(TM) format is an extension of the first format that encodes word-level scores and time alignments. It is marked by a header of the form

The hypotheses are in the format
	(score) w1 ( st: st1 et: et1 g: g1 a: a1 ) w2 ...
where words are followed by start and end times, language model and acoustic scores (bytelog-scaled), respectively. This format may also contain scores and time marks for sub-word units (phones and HMM states), in the same format as above, but with the w's denoting phone and state names. Sub-word units will have time marks that are contained in the duration of the preceding word units, and may thus be easily identified.

The third format understood by SRILM lists hypotheses in the format

	ascore lscore nwords w1 w2 w3 ...
where the first three columns contain the acoustic model log probability, the language model log probability, and the number of words in the hypothesis string, respectively. All scores are logarithms base 10. (This format must not be preceded by an ``NBestList'' header.) This format is output by the ngram -rescore and by nbest-lattice -write-nbest without the -decipher-nbest option.


ngram(1), nbest-lattice(1), segment-nbest(1), nbest-scripts(1), pfsg-scripts(1).


All these formats are somewhat ad hoc and could use a more rational design. The ``NBestList1.0'' format is particularly cumbersome because it conflates acoustic and language model scores.
A generalization to an arbitrary number of separate scores would be nice.


Manual page written by Andreas Stolcke <stolcke@speech.sri.com>.
Copyright 1999-2001 SRI International