I met some cases that when I am talking,chain model decodes right. But when I stop, it changed the result wrong(e.g. delete some words before).I wonder how to improve this. Any advice is thankful
Are you using the "big-lm-const-arpa" property? The lattice generated using the smaller LM is rescored when the decoding finishes using the "big LM", and that's sometimes changes the result noticeably.
I met some cases that when I am talking,chain model decodes right. But when I stop, it changed the result wrong(e.g. delete some words before).I wonder how to improve this. Any advice is thankful