flashlight/app/asr/tutorial/InferenceCTC.cpp - Code query

[TzurV]

Is the Trie created includes just the Unigram and a single word pronunciation that is extracted from the lexicon ?

flashlight/app/asr/tutorial/InferenceCTC.cpp: line 167

  std::shared_ptr<fl::lib::text::Trie> trie = fl::app::asr::buildTrie(
      "wrd" /* decoderType */,
      true /* useLexicon */,
      lm,
      "max" /* smearing */,
      tokenDict,
      lexicon,
      wordDict,
      wordSepIdx,
      0 /* repLabel */);
  LOG(INFO) << "[Inference tutorial for CTC] Trie is planted.";

I'm probably reading it wrongly, but I can't see were the NGram is used.

[tlikhomanenko]

Yep, it is building only unigram Trie. The ngram LM is used inside decoder itself when the leaf of trie is reached. See for example https://github.com/flashlight/flashlight/blob/master/flashlight/lib/text/decoder/LexiconDecoder.cpp#L124-L127