what does the "beam" mean in the decode model?

[新媒体gq]

may be a low question, but  it is difficult to study it, 

in the egs aishell  , there is the paramter "beam" for decode model , what does it mean or what function does it have?

eg  decode tri5a model

---------------------------

  $cmd --num-threads $num_threads JOB=1:$nj $dir/log/decode.JOB.log \

    gmm-latgen-faster$thread_string --max-active=$max_active --beam=$beam --lattice-beam=$lattice_beam \

    --acoustic-scale=$acwt --determinize-lattice=false \

    --allow-partial=true --word-symbol-table=$graphdir/words.txt \

    $adapt_model $graphdir/HCLG.fst "$pass1feats" "ark:|gzip -c > $dir/lat.tmp.JOB.gz" \

    || exit 1;

[Daniel Povey]

It determines what is pruned away in Viterbi beam search.  Larger is more accurate, smaller is faster.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/tencent_EC2C49C8EEFD8231AB5509966079BFBE7509%40qq.com.

[Hang Lyu]

In decoding, you will generate active tokens with beam search algorithem frame by frame. For each frame, the token whose cost is between "best_cost" and "best_cost + beam" will be kept, so that a temporary token list is generated. And then, the token list will be further processed by max/min-active, which should be called histgram prunning. At this point, you have got a complete active token list for a certain frame.