kaldi错误、kaldi训练错误、kaldi运行错误

　　Kaldi是最全面的语音识别开源平台，开源平台对于一些开源的总结好的数据好处理，但是对于自定义的数据，会有kaldi错误，特别是kaldi训练错误和kaldi运行错误。

　　下面就是训练错误：

　　steps/make_mfcc.sh: Succeeded creating MFCC features for train_si284_sp

　　steps/compute_cmvn_stats.sh data/train_si284_sp

　　Succeeded creating CMVN stats for train_si284_sp

　　local/nnet3/run_ivector_common.sh: fixing input data-dir to remove nonexistent features, in case some

　　.. speed-perturbed segments were too short.

　　fix_data_dir.sh: kept all 14997 utterances.

　　fix_data_dir.sh: old files are kept in data/train_si284_sp/.backup

　　local/nnet3/run_ivector_common.sh: alignments in exp/tri4b_ali_train_si284_sp appear to already exist. Please either remove them

　　... or use a later --stage option.

　　这类错误看提示就能解决，但是另外的kaldi训练错误，如下：

　　steps/train_mono.sh --boost-silence 1.25 --nj 4 --cmd run.pl data/train_si84_2kshort data/lang_nosp exp/mono0a

　　steps/train_mono.sh: Initializing monophone system.

　　feat-to-dim 'ark,s,cs:apply-cmvn --utt2spk=ark:data/train_si84_2kshort/split4/1/utt2spk scp:data/train_si84_2kshort/split4/1/cmvn.scp scp:data/train_si84_2kshort/split4/1/feats.scp ark:- | add-deltas ark:- ark:- |' -

　　apply-cmvn --utt2spk=ark:data/train_si84_2kshort/split4/1/utt2spk scp:data/train_si84_2kshort/split4/1/cmvn.scp scp:data/train_si84_2kshort/split4/1/feats.scp ark:-

　　WARNING (apply-cmvn[5.5.839~8-0c6a]:ReadScriptFile():kaldi-table:34) Error opening script file: data/train_si84_2kshort/split4/1/cmvn.scp

　　ERROR (apply-cmvn[5.5.839~8-0c6a]:RandomAccessTableReader():util/kaldi-table-inl.h:2512) Error opening RandomAccessTableReader object (rspecifier is: scp:data/train_si84_2kshort/split4/1/cmvn.scp)

　　[ Stack-Trace: ]

　　/home/server/kaldi-trunk/src/lib/libkaldi-base.so(kaldi::MessageLogger::LogMessage() const+0xb42) [0x7fe1cf528692]

　　apply-cmvn(kaldi::MessageLogger::LogAndThrow::operator=(kaldi::MessageLogger const&)+0x21) [0x55faa22a7f4f]

　　apply-cmvn(kaldi::RandomAccessTableReader<kaldi::KaldiObjectHolder<kaldi::Matrix > >::RandomAccessTableReader(std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&)+0xd3) [0x55faa22b054d]

　　apply-cmvn(kaldi::RandomAccessTableReaderMapped<kaldi::KaldiObjectHolder<kaldi::Matrix > >::RandomAccessTableReaderMapped(std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&, std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&)+0x2c) [0x55faa22b1df2]apply-cmvn(main+0x8a7) [0x55faa22a5ab1]

　　/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7fe1ce992bf7]

　　apply-cmvn(_start+0x2a) [0x55faa22a512a]

　　kaldi::KaldiFatalErroradd-deltas ark:- ark:-

　　ERROR (feat-to-dim[5.5.839~8-0c6a]:main():feat-to-dim:58) Could not read any features (empty archive?)

　　[ Stack-Trace: ]

　　/home/server/kaldi-trunk/src/lib/libkaldi-base.so(kaldi::MessageLogger::LogMessage() const+0xb42) [0x7fa2caba5692]

　　feat-to-dim(kaldi::MessageLogger::LogAndThrow::operator=(kaldi::MessageLogger const&)+0x21) [0x558d99e2b401]

　　feat-to-dim(main+0x2e9) [0x558d99e2a7f3]

　　/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7fa2ca00fbf7]

　　feat-to-dim(_start+0x2a) [0x558d99e2a42a]

　　kaldi::KaldiFatalErrorerror getting feature dimension

　　这样的就不好解决了，虽然是少文件，但是文件性质，文件出处，文件目的，这些不知道，就不好解决。开源一般是验证主功能，附加的处理比较少，对初学者不友好。

　　kaldi里有识别服务器代码，在识别时，也会有错误，比如：

　　ASSERTION_FAILED (extend-wav-with-silence[5.5.920~1-b22b8]:SubVector():matrix/kaldi-vector.h:512) Assertion failed: (static_cast(origin)+ static_cast(length) <=static_cast(t.Dim()))

　　[ Stack-Trace: ]

　　extend-wav-with-silence(kaldi::MessageLogger::LogMessage() const+0xb42) [0x55b14bbf1aca]

　　extend-wav-with-silence(kaldi::KaldiAssertFailure_(char const*, char const*, int, char const*)+0x6e) [0x55b14bbf27c6]

　　extend-wav-with-silence(kaldi::FindQuietestSegment(kaldi::Vector const&, float, kaldi::Vector*, float, float, float)+0x1a5) [0x55b14bb6b5af]

　　extend-wav-with-silence(kaldi::ExtendWaveWithSilence(kaldi::Vector const&, float, kaldi::Vector*, float, float, float)+0x3f) [0x55b14bb6b97a]

　　extend-wav-with-silence(main+0x542) [0x55b14bb6c2a5]

　　/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7) [0x7f1ee0ffeb97]

　　extend-wav-with-silence(_start+0x2a) [0x55b14bb6b32a]

　　Aborted (core dumped)

　　AMI_ES2011a_H00_FEE041_0008364_0008924 I FORGOT TO SAY I'M THE PROJECT MANAGER BUT I FIGURED JOHN KNEW THAT ALREADY

　　LOG (online2-wav-nnet3-latgen-faster[5.5.920~1-b22b8]:main():online2-wav-nnet3-latgen-faster:296) Decoded utterance AMI_ES2011a_H00_FEE041_0008364_0008924

　　WARNING (online2-wav-nnet3-latgen-faster[5.5.920~1-b22b8]:main():online2-wav-nnet3-latgen-faster:210) Did not find audio for utterance AMI_ES2011a_H00_FEE041_0009602_0009635

　　WARNING (online2-wav-nnet3-latgen-faster[5.5.920~1-b22b8]:main():online2-wav-nnet3-latgen-faster:210) Did not find audio for utterance AMI_ES2011a_H00_FEE041_0009826_0010223

　　这类断言错误，就要掌握代码细节才能解决。

　　但这些错误，如果开始遇到容易的，让开发人员都有过渡也好解决，开始就遇到难的错误，开发人员傻眼了就不好解决。这类错误需要C++和语音识别数据结构经验丰富的高手才能解决，这方面锐英源软件有经验，锐英源软件在高端GPU服务器上训练了上百套数据，几万次训练，中间经历的奇怪问题太多了。