kaldi使用online-nnet2模型时特征维度不一致问题

问

I am successfully modified the wsj examples for Verbmobil and trained models up to the local/online/run_nnet2_baseline.sh stage. Afterwards, I tried to use the model as described in Tutorial under "Example for using already-built online-nnet2 models". However, online2-wav-nnet2-latgen-faster crashes with the error message:

ERROR (online2-wav-nnet2-latgen-faster:NnetComputer():nnet-compute.cc:70) Feature dimension is 113 but network expects 13

Do I need to configure the feature dimenson somehow or am I using a wrong decoder? I am a little bit lost and would appreciate help.

我已成功修改了Verbmobil的wsj示例，并在local/online/run_nnet2_baseline.sh阶段进行了模型训练。之后，我尝试使用教程中“使用已构建的online-nnet2模型的示例”中所述的模型。但是，online2-wav-nnet2-latgen-faster崩溃并显示错误消息：

错误（online2-wav-nnet2-latgen-faster：NnetComputer（）：nnet-compute.cc:70）特征维度为113，但网络期望为13

我需要以某种方式配置特征维度还是使用错误的解码器？我有点迷糊，不胜感激。

答

You are using a config that was created for a system with iVectors while your system doesn't have iVectors (because it's _baseline.sh).
If you prepare the online-decoding directory with prepare_online_decoding.sh, as is done in run_nnet2_baseline.sh, it should set up the configs correctly, and you need to use those configs. Look at the log files decode.*.log in the decode that run_nnet2_baseline.sh does at the end, and you'll see what the decoding command should be for those models. 您正在使用为iVectors系统创建的配置， 而您的系统没有iVectors（因为它是_baseline.sh）。
如果像run_nnet2_baseline.sh一样，使用prepare_online_decoding.sh 准备联机解码目录，则它应该正确设置配置，并且需要使用这些 配置。查看 run_nnet2_baseline.sh最后执行的解码中的日志文件decode.*.log ，您将看到这些模型的解码命令应该是什么。