kaldi官方給出的解碼命令是online2-wav-nnet3-latgen-faster,它的源碼我在“基于kaldi的iOS語(yǔ)音識(shí)別(本地)+05+解碼”已經(jīng)貼出來(lái)了，下面就來(lái)詳細(xì)講解它解碼的過(guò)程，后面講解實(shí)時(shí)流解碼的時(shí)候再說(shuō)一些改進(jìn)和修改的地方。

首先我們來(lái)看它是怎么加載模型的：

    TransitionModel trans_model;
    nnet3::AmNnetSimple am_nnet;
    {
      bool binary;
      Input ki(nnet3_rxfilename, &binary);
      trans_model.Read(ki.Stream(), binary);
      am_nnet.Read(ki.Stream(), binary);
      SetBatchnormTestMode(true, &(am_nnet.GetNnet()));
      SetDropoutTestMode(true, &(am_nnet.GetNnet()));
      nnet3::CollapseModel(nnet3::CollapseModelConfig(), &(am_nnet.GetNnet()));
    }

nnet3_rxfilename就是我們給的final.mdl模型文件

Nnet3中的AmNnetSimple類(lèi)是一個(gè)標(biāo)準(zhǔn)的聲學(xué)模型類(lèi)，該類(lèi)通過(guò)調(diào)用Nnet類(lèi)進(jìn)行神經(jīng)網(wǎng)絡(luò)操作。

kaldi中的HMM模型，實(shí)際就是一個(gè)TransitionModel對(duì)象。這個(gè)對(duì)象描述了音素的HMM拓?fù)浣Y(jié)構(gòu)，并保存了pdf-id和transition-id相關(guān)的信息，并且可以進(jìn)行各種變量的轉(zhuǎn)換。

這里先不對(duì)AmNnetSimple和TransitionModel類(lèi)展開(kāi)講解，知道它是干嘛的就可以了。

nnet3::DecodableNnetSimpleLoopedInfo decodable_info(decodable_opts, &am_nnet);

此對(duì)象包含所有可解碼對(duì)象使用的預(yù)先計(jì)算的內(nèi)容。

fst::Fst<fst::StdArc> *decode_fst = ReadFstKaldiGeneric(fst_rxfilename);
fst::Fst<fst::StdArc> *decode_fst = ReadFstKaldiGeneric(fst_rxfilename);

fst::SymbolTable *word_syms = NULL;
if (word_syms_rxfilename != "")
    if (!(word_syms = fst::SymbolTable::ReadText(word_syms_rxfilename)))
    KALDI_ERR << "Could not read symbol table from file " << word_syms_rxfilename;

fst_rxfilename對(duì)應(yīng)HCLG.fst文件。word_syms_rxfilename對(duì)應(yīng)words.txt文件。

SequentialTokenVectorReader spk2utt_reader(spk2utt_rspecifier);
RandomAccessTableReader<WaveHolder> wav_reader(wav_rspecifier);
CompactLatticeWriter clat_writer(clat_wspecifier);

這里分別創(chuàng)建spk2utt_reader(說(shuō)話(huà)人+音頻)，wav_reader(待識(shí)別音頻)，clat_writer(lattice寫(xiě)入)對(duì)象。

for (; !spk2utt_reader.Done(); spk2utt_reader.Next()) {
      std::string spk = spk2utt_reader.Key();
      const std::vector<std::string> &uttlist = spk2utt_reader.Value();
      OnlineIvectorExtractorAdaptationState adaptation_state(
          feature_info.ivector_extractor_info);
      for (size_t i = 0; i < uttlist.size(); i++) {
        std::string utt = uttlist[i];
        if (!wav_reader.HasKey(utt)) {
          KALDI_WARN << "Did not find audio for utterance " << utt;
          num_err++;
          continue;
        }
      }
  ...
}

該for循環(huán)，關(guān)于說(shuō)話(huà)人音頻的內(nèi)容，我們先不關(guān)心，我們繼續(xù)其他的部分。

const WaveData &wave_data = wav_reader.Value(utt);
// get the data for channel zero (if the signal is not mono, we only
// take the first channel).
SubVector<BaseFloat> data(wave_data.Data(), 0);

根據(jù)音頻的id讀取音頻數(shù)據(jù)。

OnlineNnet2FeaturePipeline feature_pipeline(feature_info);
feature_pipeline.SetAdaptationState(adaptation_state);

OnlineSilenceWeighting silence_weighting(trans_model,
                feature_info.silence_weighting_config,
               decodable_opts.frame_subsampling_factor);

SingleUtteranceNnet3Decoder decoder(decoder_opts,
                                            trans_model,
                                            decodable_info,
                                            *decode_fst, &feature_pipeline);

OnlineNnet2FeaturePipeline 負(fù)責(zé)將神經(jīng)網(wǎng)絡(luò)的特征處理管道的各個(gè)部分組合在一起。
OnlineSilenceWeighting 負(fù)責(zé)跟蹤來(lái)自解碼器的最佳路徑回溯（有效的）并基于幀的分類(lèi)計(jì)算數(shù)據(jù)的加權(quán)在靜音（或非靜音）。
SingleUtteranceNnet3Decoder 使用神經(jīng)網(wǎng)絡(luò)的在線(xiàn)配置來(lái)解碼單個(gè)音頻。

        BaseFloat samp_freq = wave_data.SampFreq();
        int32 chunk_length;
        if (chunk_length_secs > 0) {
          chunk_length = int32(samp_freq * chunk_length_secs);
          if (chunk_length == 0) chunk_length = 1;
        } else {
          chunk_length = std::numeric_limits<int32>::max();
        }

        int32 samp_offset = 0;
        std::vector<std::pair<int32, BaseFloat> > delta_weights;

        while (samp_offset < data.Dim()) {
          int32 samp_remaining = data.Dim() - samp_offset;
          int32 num_samp = chunk_length < samp_remaining ? chunk_length
                                                         : samp_remaining;

          SubVector<BaseFloat> wave_part(data, samp_offset, num_samp);
          feature_pipeline.AcceptWaveform(samp_freq, wave_part);

          samp_offset += num_samp;
          decoding_timer.WaitUntil(samp_offset / samp_freq);
          if (samp_offset == data.Dim()) {
            // no more input. flush out last frames
            feature_pipeline.InputFinished();
          }

          if (silence_weighting.Active() &&
              feature_pipeline.IvectorFeature() != NULL) {
            silence_weighting.ComputeCurrentTraceback(decoder.Decoder());
            silence_weighting.GetDeltaWeights(feature_pipeline.NumFramesReady(),
                                              &delta_weights);
            feature_pipeline.IvectorFeature()->UpdateFrameWeights(delta_weights);
          }

          decoder.AdvanceDecoding();

          if (do_endpointing && decoder.EndpointDetected(endpoint_opts)) {
            break;
          }
        }

將音頻一段一段給feature_pipeline同時(shí)開(kāi)啟解碼decoder.AdvanceDecoding();

decoder.FinalizeDecoding();

完成解碼