[egs] LT-LM recipe for librispeech (#4590)

Co-authored-by: Anton Mitrofanov <[email protected]>

Author: medbar

egs/librispeech/s5/fairseq_ltlm/README.md

@@ -0,0 +1,12 @@
+# LT-LM: a novel non-autoregressive language model for single-shot lattice rescoring
+[Paper](https://arxiv.org/pdf/2104.02526.pdf)
+
+## Setup:
+`cd fairseq_ltlm && setup.sh`
+## run:
+* put slurm.conf to conf/
+* modify fairseq\_ltlm/recipes/config.sh if needed
+* `bash fairseq\_ltlm/recipes/run.sh`
+## evaluate:
+For evaluation, you can 
+run fairseq\_ltlm/recipes/run\_5\_eval.sh (see run.sh) or use fairseq\_ltlm/ltlm/eval.py directly.

egs/librispeech/s5/fairseq_ltlm/kaldi_utils/Makefile

@@ -0,0 +1,22 @@
+kaldi=../../../../..
+all:
+
+include $(kaldi)/src/kaldi.mk
+
+EXTRA_CXXFLAGS += -Wno-sign-compare
+EXTRA_CXXFLAGS += -I$(kaldi)/src
+BINFILES = latgen-faster-mapped-fake-am
+
+OBJFILES =
+
+TESTFILES =
+
+ADDLIBS = $(kaldi)/src/decoder/kaldi-decoder.a $(kaldi)/src/lat/kaldi-lat.a $(kaldi)/src/lm/kaldi-lm.a \
+          $(kaldi)/src/fstext/kaldi-fstext.a $(kaldi)/src/hmm/kaldi-hmm.a \
+          $(kaldi)/src/transform/kaldi-transform.a $(kaldi)/src/gmm/kaldi-gmm.a \
+          $(kaldi)/src/tree/kaldi-tree.a $(kaldi)/src/util/kaldi-util.a $(kaldi)/src/matrix/kaldi-matrix.a \
+          $(kaldi)/src/base/kaldi-base.a
+
+
+include $(kaldi)/src/makefiles/default_rules.mk
+

egs/librispeech/s5/fairseq_ltlm/kaldi_utils/latgen-faster-mapped-fake-am.cc

@@ -0,0 +1,213 @@
+
+// Copyright (c) 2021, Speech Technology Center Ltd. All rights reserved.
+// Anton Mitrofanov
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// It is latgen-faster-mapped adopted to fake lattice generation
+
+#include <chrono>
+
+#include "base/kaldi-common.h"
+#include "base/timer.h"
+#include "decoder/decodable-matrix.h"
+#include "decoder/decoder-wrappers.h"
+#include "fstext/fstext-lib.h"
+#include "hmm/transition-model.h"
+#include "tree/context-dep.h"
+#include "util/common-utils.h"
+
+using namespace kaldi;
+typedef kaldi::int32 int32;
+using fst::Fst;
+using fst::StdArc;
+using fst::SymbolTable;
+
+int main(int argc, char *argv[]) {
+  try {
+    const char *usage =
+        "Generate lattices, reading emulating am as matrices\n"
+        " (model is needed only for the integer mappings in its "
+        "transition-model)\n"
+        "Usage: latgen-faster-mapped-fake-am [options] trans-model-in fst-in "
+        "fam-rspecifier ali_rspecifier"
+        " lattice-wspecifier [ words-wspecifier [alignments-wspecifier] ]\n";
+    ParseOptions po(usage);
+    Timer timer;
+    bool allow_partial = false;
+    BaseFloat acoustic_scale = 0.1;
+    LatticeFasterDecoderConfig config;
+
+    std::string word_syms_filename;
+    config.Register(&po);
+    po.Register("acoustic-scale", &acoustic_scale,
+                "Scaling factor for acoustic likelihoods");
+
+    po.Register("word-symbol-table", &word_syms_filename,
+                "Symbol table for words [for debug output]");
+    po.Register("allow-partial", &allow_partial,
+                "If true, produce output even if end state was not reached.");
+
+    po.Read(argc, argv);
+
+    if (po.NumArgs() < 5 || po.NumArgs() > 7) {
+      po.PrintUsage();
+      exit(1);
+    }
+
+    std::string model_in_filename = po.GetArg(1), fst_in_str = po.GetArg(2),
+                fam_rspecifier = po.GetArg(3), ali_rspecifier = po.GetArg(4),
+                lattice_wspecifier = po.GetArg(5),
+                words_wspecifier = po.GetOptArg(6),
+                alignment_wspecifier = po.GetOptArg(7);
+
+    TransitionModel trans_model;
+    ReadKaldiObject(model_in_filename, &trans_model);
+
+    bool determinize = config.determinize_lattice;
+    CompactLatticeWriter compact_lattice_writer;
+    LatticeWriter lattice_writer;
+    if (!(determinize ? compact_lattice_writer.Open(lattice_wspecifier)
+                      : lattice_writer.Open(lattice_wspecifier)))
+      KALDI_ERR << "Could not open table for writing lattices: "
+                << lattice_wspecifier;
+
+    Int32VectorWriter words_writer(words_wspecifier);
+
+    Int32VectorWriter alignment_writer(alignment_wspecifier);
+
+    fst::SymbolTable *word_syms = NULL;
+    if (word_syms_filename != "")
+      if (!(word_syms = fst::SymbolTable::ReadText(word_syms_filename)))
+        KALDI_ERR << "Could not read symbol table from file "
+                  << word_syms_filename;
+
+    double tot_like = 0.0;
+    kaldi::int64 frame_count = 0;
+    int num_success = 0, num_fail = 0;
+
+    // Reading Fake acoustic model form ark file
+    KALDI_LOG << "Loading Fake Acoustic Model";
+    SequentialBaseFloatMatrixReader fam_model_read(fam_rspecifier);
+    std::string fam_model_key = fam_model_read.Key();
+    Matrix<BaseFloat> fam_model(fam_model_read.Value());
+    KALDI_LOG << "Apply log.";
+    fam_model.ApplyLog();
+
+    if (fam_model_key != "fam_model") {
+      KALDI_ERR << fam_rspecifier << " - Wrong fam_model.";
+      po.PrintUsage();
+      exit(1);
+    }
+    KALDI_LOG << "Fake Acoustic is loaded. Shape is (" << fam_model.NumRows()
+              << ", " << fam_model.NumCols() << ")";
+
+    SequentialInt32VectorReader ali_reader(ali_rspecifier);
+    if (ClassifyRspecifier(fst_in_str, NULL, NULL) == kNoRspecifier) {
+      // SequentialBaseFloatMatrixReader loglike_reader(feature_rspecifier);
+      // Input FST is just one FST, not a table of FSTs.
+      Fst<StdArc> *decode_fst = fst::ReadFstKaldiGeneric(fst_in_str);
+      timer.Reset();
+
+      {
+        LatticeFasterDecoder decoder(*decode_fst, config);
+        for (; !ali_reader.Done(); ali_reader.Next()) {
+          std::string utt = ali_reader.Key();
+          std::vector<int32> ali(ali_reader.Value());
+          KALDI_LOG << "Process " << utt << ". " << ali.size() << " frames";
+          ali_reader.FreeCurrent();
+          if (ali.size() == 0) {
+            KALDI_WARN << "Zero-length utterance: " << utt;
+            num_fail++;
+            continue;
+          }
+          // Inference fake AM
+          kaldi::Matrix<BaseFloat> loglikes(ali.size(), fam_model.NumRows());
+          loglikes.SetZero();
+          for (int i = 0; i < ali.size(); i++) {
+            int32 pdf_id = ali[i];
+            loglikes.CopyRowFromVec(fam_model.Row(pdf_id), i);
+            SubVector<BaseFloat> row(loglikes, i);
+          }
+          // end
+          DecodableMatrixScaledMapped decodable(trans_model, loglikes,
+                                                acoustic_scale);
+
+          double like;
+          if (DecodeUtteranceLatticeFaster(
+                  decoder, decodable, trans_model, word_syms, utt,
+                  acoustic_scale, determinize, allow_partial, &alignment_writer,
+                  &words_writer, &compact_lattice_writer, &lattice_writer,
+                  &like)) {
+            tot_like += like;
+            frame_count += loglikes.NumRows();
+            num_success++;
+          } else
+            num_fail++;
+        }
+      }
+      delete decode_fst;  // delete this only after decoder goes out of scope.
+    } else {              // We have different FSTs for different utterances.
+      KALDI_LOG << "FSTs not implemented yet.";
+      exit(1);
+      //      SequentialTableReader<fst::VectorFstHolder>
+      //      fst_reader(fst_in_str); RandomAccessBaseFloatMatrixReader
+      //      loglike_reader(feature_rspecifier); for (; !fst_reader.Done();
+      //      fst_reader.Next()) {
+      //        std::string utt = fst_reader.Key();
+      //        if (!loglike_reader.HasKey(utt)) {
+      //          KALDI_WARN << "Not decoding utterance " << utt
+      //                     << " because no loglikes available.";
+      //          num_fail++;
+      //          continue;
+      //        }
+      //        const Matrix<BaseFloat> &loglikes = loglike_reader.Value(utt);
+      //        if (loglikes.NumRows() == 0) {
+      //          KALDI_WARN << "Zero-length utterance: " << utt;
+      //          num_fail++;
+      //          continue;
+      //        }
+      //        LatticeFasterDecoder decoder(fst_reader.Value(), config);
+      //        DecodableMatrixScaledMapped decodable(trans_model, loglikes,
+      //        acoustic_scale); double like; if (DecodeUtteranceLatticeFaster(
+      //                decoder, decodable, trans_model, word_syms, utt,
+      //                acoustic_scale, determinize, allow_partial,
+      //                &alignment_writer, &words_writer,
+      //                &compact_lattice_writer, &lattice_writer, &like)) {
+      //          tot_like += like;
+      //          frame_count += loglikes.NumRows();
+      //          num_success++;
+      //        } else num_fail++;
+      //      }
+    }
+
+    double elapsed = timer.Elapsed();
+    KALDI_LOG << "Time taken " << elapsed
+              << "s: real-time factor assuming 100 frames/sec is "
+              << (elapsed * 100.0 / frame_count);
+    KALDI_LOG << "Done " << num_success << " utterances, failed for "
+              << num_fail;
+    KALDI_LOG << "Overall log-likelihood per frame is "
+              << (tot_like / frame_count) << " over " << frame_count
+              << " frames.";
+
+    delete word_syms;
+    if (num_success != 0)
+      return 0;
+    else
+      return 1;
+  } catch (const std::exception &e) {
+    std::cerr << e.what();
+    return -1;
+  }
+}

egs/librispeech/s5/fairseq_ltlm/ltlm/Tokenizer.py

@@ -0,0 +1,93 @@
+# Copyright 2021 STC-Innovation LTD (Author: Anton Mitrofanov)
+import argparse
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class WordTokenizer:
+    """ A words.txt mapping"""
+    @staticmethod
+    def add_args(parser: argparse.ArgumentParser):
+        parser.add_argument('--tokenizer_fn', type=str, required=True,
+                            help='Tokenizer fname( words.txt)')
+        #parser.add_argument('--unk', default='<UNK>', help="Unk word") # fairseq bug
+
+    @staticmethod
+    def build_from_args(args):
+        kwargs = {"fname": args.tokenizer_fn}
+                  #'unk': args.unk_word}
+
+        return WordTokenizer(**kwargs)
+
+    def __init__(self, fname, unk_word='<UNK>'):
+        logger.info(f'Loading WordTokenizer {fname}')
+        with open(fname, 'r', encoding='utf-8') as f:
+            self.word2id = {w: int(i) for w, i in map(str.split, f.readlines())}
+        self.id2word = ['']*(max(self.word2id.values()) + 1)
+        self.unk=unk_word
+        if self.unk not in self.word2id:
+            if self.unk.lower() in self.word2id:
+                self.unk=self.unk.lower()
+            else:
+                raise f"unk word {unk_word} not in {fname}"
+        for w, i in self.word2id.items():
+            self.id2word[i] = w
+        assert self.word2id['<eps>'] == 0 and \
+            '<s>' in self.word2id.keys() and \
+            '</s>' in self.word2id.keys(), RuntimeError("<esp>!=0")
+
+        self.real_words_ids = [i for w, i in self.word2id.items() \
+                               if w.find('<') == w.find('>') == w.find('#') == w.find('!') == w.find('[') == w.find(']') == -1 and \
+                               not w.endswith('-') and not w.startswith("-") ]
+
+        logger.info(f'WordTokenizer {fname} loaded. Vocab size {len(self)}.')
+        self.disambig_word_ids = [i for w, i in self.word2id.items() \
+                                    if (w != "<s>" and w != "</s>") and (
+                                        w.find('<') != -1 or
+                                        w.find('>') != -1 or
+                                        w.find('#') != -1 or
+                                        w.find('!') != -1 or 
+                                        w.find('[') != -1 or 
+                                        w.find(']') != -1 or 
+                                        w.endswith('-')   or 
+                                        w.startswith('-'))]
+        logger.info(f"WordTokenizer Disambig ids: {self.disambig_word_ids}")
+        logger.info(f"WordTokenizer Disambig words: {[ self.id2word[i] for i in self.disambig_word_ids]}")
+
+    def encode(self, text, bos=False, eos=False):
+        return [
+            ([self.get_bos_id()] if bos else []) +
+             [self.word2id[w] if w in self.word2id.keys() else self.word2id[self.unk] for w in line.split()] +
+            ([self.get_eos_id()] if eos else []) for line in text]
+
+    def decode(self, text_ids):
+        return [[self.id2word[i] for i in line_ids] for line_ids in text_ids]
+
+    def __len__(self):
+        return len(self.id2word)
+
+    def get_real_words_ids(self):
+        return self.real_words_ids
+
+    def get_disambig_words_ids(self):
+        return self.disambig_word_ids
+
+    def get_bos_id(self):
+        return self.word2id["<s>"]
+
+    def get_eos_id(self):
+        return self.word2id["</s>"]
+
+    def get_unk_id(self):
+        return self.word2id[self.unk]
+
+    def pad(self):
+        return self.word2id['<eps>']
+
+    def print_lat(self, lat, print_word_id=False, p=None):
+        for i, arc in enumerate(lat):
+            out_str = f"{arc[1]} {arc[2]} {arc[0] if print_word_id else self.id2word[arc[0]]}"
+            if p is not None:
+                out_str += f" {p[i]}"
+            print(out_str)

egs/librispeech/s5/fairseq_ltlm/ltlm/__init__.py

@@ -0,0 +1,5 @@
+# Copyright 2021 STC-Innovation LTD (Author: Anton Mitrofanov)
+from .criterions.bce_loss import BCECriterion
+from .datasets import LatsOracleAlignDataSet
+from .models import LTLM
+from .tasks.rescoring_task import RescoringTask