Implement recipe for Fluent Speech Commands dataset (#1469)

---------

Signed-off-by: Xinyuan Li <xli257@c13.clsp.jhu.edu>

egs/fluent_speech_commands/SLU/README.md

@@ -0,0 +1,9 @@
+## Fluent Speech Commands recipe
+
+This is a recipe for the Fluent Speech Commands dataset, a speech dataset which transcribes short utterances (such as "turn the lights on in the kitchen") into action frames (such as {"action": "activate", "object": "lights", "location": "kitchen"}). The training set contains 23,132 utterances, whereas the test set contains 3793 utterances. 
+
+Dataset Paper link: <https://paperswithcode.com/dataset/fluent-speech-commands>
+
+cd icefall/egs/fluent_speech_commands/
+Training: python transducer/train.py
+Decoding: python transducer/decode.py

egs/fluent_speech_commands/SLU/local/compile_hlg.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str) -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    logging.info("Loading G.fst.txt")
+    with open(lang_dir / "G.fst.txt") as f:
+        G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    # LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # see https://github.com/k2-fsa/k2/pull/1140
+    labels = LG.labels
+    labels[labels >= first_token_disambig_id] = 0
+    LG.labels = labels
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lang_dir)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/fluent_speech_commands/SLU/local/compute_fbank_slu.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+
+"""
+This file computes fbank features of the Fluent Speech Commands dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or it wastes a
+# lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_slu(manifest_dir, fbanks_dir):
+    src_dir = Path(manifest_dir)
+    output_dir = Path(fbanks_dir)
+
+    # This dataset is rather small, so we use only one job
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 23
+
+    dataset_parts = (
+        "train",
+        "valid",
+        "test",
+    )
+    prefix = "slu"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(sampling_rate=16000, num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_file = output_dir / f"{prefix}_cuts_{partition}.{suffix}"
+            if cuts_file.is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 1,  # use one job
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(cuts_file)
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument("manifest_dir")
+parser.add_argument("fbanks_dir")
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    args = parser.parse_args()
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_slu(args.manifest_dir, args.fbanks_dir)

egs/fluent_speech_commands/SLU/local/generate_lexicon.py

@@ -0,0 +1,59 @@
+import argparse
+
+import pandas
+from tqdm import tqdm
+
+
+def generate_lexicon(corpus_dir, lm_dir):
+    data = pandas.read_csv(
+        str(corpus_dir) + "/data/train_data.csv", index_col=0, header=0
+    )
+    vocab_transcript = set()
+    vocab_frames = set()
+    transcripts = data["transcription"].tolist()
+    frames = list(
+        i
+        for i in zip(
+            data["action"].tolist(), data["object"].tolist(), data["location"].tolist()
+        )
+    )
+
+    for transcript in tqdm(transcripts):
+        for word in transcript.split():
+            vocab_transcript.add(word)
+
+    for frame in tqdm(frames):
+        for word in frame:
+            vocab_frames.add("_".join(word.split()))
+
+    with open(lm_dir + "/words_transcript.txt", "w") as lexicon_transcript_file:
+        lexicon_transcript_file.write("<UNK> 1" + "\n")
+        lexicon_transcript_file.write("<s> 2" + "\n")
+        lexicon_transcript_file.write("</s> 0" + "\n")
+        id = 3
+        for vocab in vocab_transcript:
+            lexicon_transcript_file.write(vocab + " " + str(id) + "\n")
+            id += 1
+
+    with open(lm_dir + "/words_frames.txt", "w") as lexicon_frames_file:
+        lexicon_frames_file.write("<UNK> 1" + "\n")
+        lexicon_frames_file.write("<s> 2" + "\n")
+        lexicon_frames_file.write("</s> 0" + "\n")
+        id = 3
+        for vocab in vocab_frames:
+            lexicon_frames_file.write(vocab + " " + str(id) + "\n")
+            id += 1
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument("corpus_dir")
+parser.add_argument("lm_dir")
+
+
+def main():
+    args = parser.parse_args()
+
+    generate_lexicon(args.corpus_dir, args.lm_dir)
+
+
+main()

egs/fluent_speech_commands/SLU/local/prepare_lang.py

@@ -0,0 +1,371 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "!SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    # assert token2id["<eps>"] == 0
+    # assert word2id["<eps>"] == 0
+
+    eps = 0
+    sil_token = word2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [word2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = word2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument("lm_dir")
+
+
+def main():
+    args = parser.parse_args()
+
+    out_dir = Path(args.lm_dir)
+    lexicon_filenames = [out_dir / "words_frames.txt", out_dir / "words_transcript.txt"]
+    names = ["frames", "transcript"]
+    sil_token = "!SIL"
+    sil_prob = 0.5
+
+    for name, lexicon_filename in zip(names, lexicon_filenames):
+        lexicon = read_lexicon(lexicon_filename)
+        tokens = get_words(lexicon)
+        words = get_words(lexicon)
+        new_lexicon = []
+        for lexicon_item in lexicon:
+            new_lexicon.append((lexicon_item[0], [lexicon_item[0]]))
+        lexicon = new_lexicon
+
+        lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+        for i in range(max_disambig + 1):
+            disambig = f"#{i}"
+            assert disambig not in tokens
+            tokens.append(f"#{i}")
+
+        tokens = ["<eps>"] + tokens
+        words = ["eps"] + words + ["#0", "!SIL"]
+
+        token2id = generate_id_map(tokens)
+        word2id = generate_id_map(words)
+
+        write_mapping(out_dir / ("tokens_" + name + ".txt"), token2id)
+        write_mapping(out_dir / ("words_" + name + ".txt"), word2id)
+        write_lexicon(out_dir / ("lexicon_disambig_" + name + ".txt"), lexicon_disambig)
+
+        L = lexicon_to_fst(
+            lexicon,
+            token2id=word2id,
+            word2id=word2id,
+            sil_token=sil_token,
+            sil_prob=sil_prob,
+        )
+
+        L_disambig = lexicon_to_fst(
+            lexicon_disambig,
+            token2id=word2id,
+            word2id=word2id,
+            sil_token=sil_token,
+            sil_prob=sil_prob,
+            need_self_loops=True,
+        )
+        torch.save(L.as_dict(), out_dir / ("L_" + name + ".pt"))
+        torch.save(L_disambig.as_dict(), out_dir / ("L_disambig_" + name + ".pt"))
+
+        if False:
+            # Just for debugging, will remove it
+            L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+            L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+            L_disambig.labels_sym = L.labels_sym
+            L_disambig.aux_labels_sym = L.aux_labels_sym
+            L.draw(out_dir / "L.png", title="L")
+            L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+main()

egs/fluent_speech_commands/SLU/prepare.sh

@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+stage=1
+stop_stage=5
+
+data_dir=path/to/fluent/speech/commands
+target_root_dir=data/
+
+lang_dir=${target_root_dir}/lang_phone
+lm_dir=${target_root_dir}/lm
+manifest_dir=${target_root_dir}/manifests
+fbanks_dir=${target_root_dir}/fbanks
+
+. shared/parse_options.sh || exit 1
+
+mkdir -p $lang_dir
+mkdir -p $lm_dir
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "data_dir: $data_dir"
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare slu manifest"
+  mkdir -p $manifest_dir
+  lhotse prepare slu $data_dir $manifest_dir
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute fbank for SLU"
+  mkdir -p $fbanks_dir
+  python ./local/compute_fbank_slu.py $manifest_dir $fbanks_dir
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare lang"
+  # NOTE: "<UNK> SIL" is added for implementation convenience
+  # as the graph compiler code requires that there is a OOV word
+  # in the lexicon.
+  python ./local/generate_lexicon.py $data_dir $lm_dir
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Train LM"
+  # We use a unigram G
+  ./shared/make_kn_lm.py \
+    -ngram-order 1 \
+    -text $lm_dir/words_transcript.txt \
+    -lm $lm_dir/G_transcript.arpa
+
+  ./shared/make_kn_lm.py \
+    -ngram-order 1 \
+    -text $lm_dir/words_frames.txt \
+    -lm $lm_dir/G_frames.arpa
+
+  python ./local/prepare_lang.py $lm_dir
+
+  if [ ! -f $lm_dir/G_transcript.fst.txt ]; then
+    python -m kaldilm \
+      --read-symbol-table="$lm_dir/words_transcript.txt" \
+      $lm_dir/G_transcript.arpa > $lm_dir/G_transcript.fst.txt
+  fi
+
+  if [ ! -f $lm_dir/G_frames.fst.txt ]; then
+    python -m kaldilm \
+      --read-symbol-table="$lm_dir/words_frames.txt" \
+      $lm_dir/G_frames.arpa > $lm_dir/G_frames.fst.txt
+  fi
+
+  mkdir -p $lm_dir/frames
+  mkdir -p $lm_dir/transcript
+
+  chmod -R +777 .
+
+  for i in G_frames.arpa G_frames.fst.txt L_disambig_frames.pt L_frames.pt lexicon_disambig_frames.txt tokens_frames.txt words_frames.txt;
+  do
+    j=${i//"_frames"/}
+    mv "$lm_dir/$i" $lm_dir/frames/$j
+  done
+
+  for i in G_transcript.arpa G_transcript.fst.txt L_disambig_transcript.pt L_transcript.pt lexicon_disambig_transcript.txt tokens_transcript.txt words_transcript.txt;
+  do
+    j=${i//"_transcript"/}
+    mv "$lm_dir/$i" $lm_dir/transcript/$j
+  done
+fi
+
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compile HLG"
+  ./local/compile_hlg.py --lang-dir $lm_dir/frames
+  ./local/compile_hlg.py --lang-dir $lm_dir/transcript
+
+fi

egs/fluent_speech_commands/SLU/shared

@@ -0,0 +1 @@
+../../icefall/shared/

egs/fluent_speech_commands/SLU/transducer/beam_search.py

@@ -0,0 +1,71 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+from typing import List
+
+import torch
+from transducer.model import Transducer
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, id2word: dict
+) -> List[str]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    device = model.device
+
+    sos = torch.tensor([blank_id], device=device).reshape(1, 1)
+    decoder_out, (h, c) = model.decoder(sos)
+    T = encoder_out.size(1)
+    t = 0
+    hyp = []
+    max_u = 1000  # terminate after this number of steps
+    u = 0
+
+    while t < T and u < max_u:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(current_encoder_out, decoder_out)
+
+        log_prob = logits.log_softmax(dim=-1)
+        # log_prob is (N, 1, 1)
+        # TODO: Use logits.argmax()
+        y = log_prob.argmax()
+        if y != blank_id:
+            hyp.append(y.item())
+            y = y.reshape(1, 1)
+            decoder_out, (h, c) = model.decoder(y, (h, c))
+            u += 1
+        else:
+            t += 1
+    # id2word = {1: "YES", 2: "NO"}
+
+    hyp = [id2word[i] for i in hyp]
+
+    return hyp

egs/fluent_speech_commands/SLU/transducer/conformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/conformer.py

egs/fluent_speech_commands/SLU/transducer/decode.py

@@ -0,0 +1,346 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+from transducer.beam_search import greedy_search
+from transducer.conformer import Conformer
+from transducer.decoder import Decoder
+from transducer.joiner import Joiner
+from transducer.model import Transducer
+from transducer.slu_datamodule import SluDataModule
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_id2word(params):
+    id2word = {}
+
+    # 0 is blank
+    id = 1
+    try:
+        with open(Path(params.lang_dir) / "lexicon_disambig.txt") as lexicon_file:
+            for line in lexicon_file:
+                if len(line.strip()) > 0:
+                    id2word[id] = line.split()[0]
+                    id += 1
+    except:
+        pass
+
+    return id2word
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=6,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="Directory from which to load the checkpoints",
+    )
+    parser.add_argument("--lang-dir", type=str, default="data/lm/frames")
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 23,
+            "lang_dir": Path("data/lm/frames"),
+            # encoder/decoder params
+            "vocab_size": 3,  # blank, yes, no
+            "blank_id": 0,
+            "embedding_dim": 32,
+            "hidden_dim": 16,
+            "num_decoder_layers": 4,
+        }
+    )
+
+    vocab_size = 1
+    with open(params.lang_dir / "lexicon_disambig.txt") as lexicon_file:
+        for line in lexicon_file:
+            if (
+                len(line.strip()) > 0
+            ):  # and '<UNK>' not in line and '<s>' not in line and '</s>' not in line:
+                vocab_size += 1
+    params.vocab_size = vocab_size
+
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict, model: nn.Module, batch: dict, id2word: dict
+) -> List[List[int]]:
+    """Decode one batch and return the result in a list-of-list.
+    Each sub list contains the word IDs for an utterance in the batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding.
+        - params.method is "nbest", it uses nbest decoding.
+
+      model:
+        The neural model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+        (https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py)
+    Returns:
+      Return the decoding result. `len(ans)` == batch size.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+    feature_lens = batch["supervisions"]["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        hyp = greedy_search(model=model, encoder_out=encoder_out_i, id2word=id2word)
+        hyps.append(hyp)
+
+    #  hyps = [[word_table[i] for i in ids] for ids in hyps]
+    return hyps
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+) -> List[Tuple[List[int], List[int]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+    Returns:
+      Return a tuple contains two elements (ref_text, hyp_text):
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    id2word = get_id2word(params)
+
+    results = []
+    for batch_idx, batch in enumerate(dl):
+        texts = [
+            " ".join(a.supervisions[0].custom["frames"])
+            for a in batch["supervisions"]["cut"]
+        ]
+        texts = [
+            "<s> " + a.replace("change language", "change_language") + " </s>"
+            for a in texts
+        ]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps = decode_one_batch(
+            params=params, model=model, batch=batch, id2word=id2word
+        )
+
+        this_batch = []
+        assert len(hyps) == len(texts)
+        for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+            ref_words = ref_text.split()
+            this_batch.append((cut_id, ref_words, hyp_words))
+
+        results.extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    exp_dir: Path,
+    test_set_name: str,
+    results: List[Tuple[List[int], List[int]]],
+) -> None:
+    """Save results to `exp_dir`.
+    Args:
+      exp_dir:
+        The output directory. This function create the following files inside
+        this directory:
+
+            - recogs-{test_set_name}.text
+
+                It contains the reference and hypothesis results, like below::
+
+                    ref=['NO', 'NO', 'NO', 'YES', 'NO', 'NO', 'NO', 'YES']
+                    hyp=['NO', 'NO', 'NO', 'YES', 'NO', 'NO', 'NO', 'YES']
+                    ref=['NO', 'NO', 'YES', 'NO', 'YES', 'NO', 'NO', 'YES']
+                    hyp=['NO', 'NO', 'YES', 'NO', 'YES', 'NO', 'NO', 'YES']
+
+            - errs-{test_set_name}.txt
+
+                It contains the detailed WER.
+      test_set_name:
+        The name of the test set, which will be part of the result filename.
+      results:
+        A list of tuples, each of which contains (ref_words, hyp_words).
+    Returns:
+      Return None.
+    """
+    recog_path = exp_dir / f"recogs-{test_set_name}.txt"
+    results = sorted(results)
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    # The following prints out WERs, per-word error statistics and aligned
+    # ref/hyp pairs.
+    errs_filename = exp_dir / f"errs-{test_set_name}.txt"
+    with open(errs_filename, "w") as f:
+        write_error_stats(f, f"{test_set_name}", results)
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+
+def get_transducer_model(params: AttributeDict):
+    # encoder = Tdnn(
+    #     num_features=params.feature_dim,
+    #     output_dim=params.hidden_dim,
+    # )
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.hidden_dim,
+    )
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.hidden_dim,
+        embedding_dropout=0.4,
+        rnn_dropout=0.4,
+    )
+    joiner = Joiner(input_dim=params.hidden_dim, output_dim=params.vocab_size)
+    transducer = Transducer(encoder=encoder, decoder=decoder, joiner=joiner)
+    return transducer
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    SluDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+    params["env_info"] = get_env_info()
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    slu = SluDataModule(args)
+    test_dl = slu.test_dataloaders()
+    results = decode_dataset(
+        dl=test_dl,
+        params=params,
+        model=model,
+    )
+
+    test_set_name = str(args.feature_dir).split("/")[-2]
+    save_results(exp_dir=params.exp_dir, test_set_name=test_set_name, results=results)
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/fluent_speech_commands/SLU/transducer/decoder.py

@@ -0,0 +1 @@
+../../../yesno/ASR/transducer/decoder.py

egs/fluent_speech_commands/SLU/transducer/encoder_interface.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py

egs/fluent_speech_commands/SLU/transducer/joiner.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer/joiner.py

egs/fluent_speech_commands/SLU/transducer/model.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer/model.py

egs/fluent_speech_commands/SLU/transducer/slu_datamodule.py

@@ -0,0 +1,289 @@
+# Copyright      2021  Piotr Żelasko
+#                2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import List
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.dataset.datamodule import DataModule
+from icefall.utils import str2bool
+
+
+class SluDataModule(DataModule):
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    """
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        super().add_arguments(parser)
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--feature-dir",
+            type=Path,
+            default=Path("data/fbanks"),
+            help="Path to directory with train/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=30.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=False,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=10,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+    def train_dataloaders(self) -> DataLoader:
+        logging.info("About to get train cuts")
+        cuts_train = self.train_cuts()
+
+        logging.info("About to create train dataset")
+        transforms = []
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(
+                    FbankConfig(sampling_rate=8000, num_mel_bins=23)
+                ),
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self) -> DataLoader:
+        logging.info("About to get valid cuts")
+        cuts_valid = self.valid_cuts()
+
+        logging.debug("About to create valid dataset")
+        valid = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=23)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create valid dataloader")
+        valid_dl = DataLoader(
+            valid,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+        )
+        return valid_dl
+
+    def test_dataloaders(self) -> DataLoader:
+        logging.info("About to get test cuts")
+        cuts_test = self.test_cuts()
+
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=23)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts_test,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=True,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest_lazy(
+            self.args.feature_dir / "slu_cuts_train.jsonl.gz"
+        )
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> List[CutSet]:
+        logging.info("About to get valid cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.feature_dir / "slu_cuts_valid.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        cuts_test = load_manifest_lazy(self.args.feature_dir / "slu_cuts_test.jsonl.gz")
+        return cuts_test

egs/fluent_speech_commands/SLU/transducer/subsampling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/subsampling.py

egs/fluent_speech_commands/SLU/transducer/test_conformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer/test_conformer.py

egs/fluent_speech_commands/SLU/transducer/test_decoder.py

@@ -0,0 +1 @@
+../../../yesno/ASR/transducer/test_decoder.py

egs/fluent_speech_commands/SLU/transducer/test_joiner.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer/test_joiner.py

egs/fluent_speech_commands/SLU/transducer/test_transducer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer/test_transducer.py

egs/fluent_speech_commands/SLU/transducer/train.py

@@ -0,0 +1,625 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from lhotse.utils import fix_random_seed
+from slu_datamodule import SluDataModule
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from transducer.conformer import Conformer
+
+# from torch.utils.tensorboard import SummaryWriter
+from transducer.decoder import Decoder
+from transducer.joiner import Joiner
+from transducer.model import Transducer
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_word2id(params):
+    word2id = {}
+
+    # 0 is blank
+    id = 1
+    with open(Path(params.lang_dir) / "lexicon_disambig.txt") as lexicon_file:
+        for line in lexicon_file:
+            if len(line.strip()) > 0:
+                word2id[line.split()[0]] = id
+                id += 1
+
+    return word2id
+
+
+def get_labels(texts: List[str], word2id) -> k2.RaggedTensor:
+    """
+    Args:
+      texts:
+        A list of transcripts.
+    Returns:
+      Return a ragged tensor containing the corresponding word ID.
+    """
+    # blank is 0
+    word_ids = []
+    for t in texts:
+        words = t.split()
+        ids = [word2id[w] for w in words]
+        word_ids.append(ids)
+
+    return k2.RaggedTensor(word_ids)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=7,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer/exp",
+        help="Directory to save results",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument("--lang-dir", type=str, default="data/lm/frames")
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - start_epoch:  If it is not zero, load checkpoint `start_epoch-1`
+                        and continue training from that checkpoint.
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+
+    """
+    params = AttributeDict(
+        {
+            "lr": 1e-4,
+            "feature_dim": 23,
+            "weight_decay": 1e-6,
+            "start_epoch": 0,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 100,
+            "reset_interval": 20,
+            "valid_interval": 3000,
+            "exp_dir": Path("transducer/exp"),
+            "lang_dir": Path("data/lm/frames"),
+            # encoder/decoder params
+            "vocab_size": 3,  # blank, yes, no
+            "blank_id": 0,
+            "embedding_dim": 32,
+            "hidden_dim": 16,
+            "num_decoder_layers": 4,
+        }
+    )
+
+    vocab_size = 1
+    with open(Path(params.lang_dir) / "lexicon_disambig.txt") as lexicon_file:
+        for line in lexicon_file:
+            if (
+                len(line.strip()) > 0
+            ):  # and '<UNK>' not in line and '<s>' not in line and '</s>' not in line:
+                vocab_size += 1
+    params.vocab_size = vocab_size
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict, model: nn.Module, batch: dict, is_training: bool, word2ids
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Tdnn in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    feature_lens = batch["supervisions"]["num_frames"].to(device)
+
+    texts = [
+        " ".join(a.supervisions[0].custom["frames"])
+        for a in batch["supervisions"]["cut"]
+    ]
+    texts = [
+        "<s> " + a.replace("change language", "change_language") + " </s>"
+        for a in texts
+    ]
+    labels = get_labels(texts, word2ids).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(x=feature, x_lens=feature_lens, y=labels)
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = feature.size(0)
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    valid_dl: torch.utils.data.DataLoader,
+    word2ids,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            is_training=False,
+            word2ids=word2ids,
+        )
+        assert loss.requires_grad is False
+
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    word2ids,
+    tb_writer: None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params, model=model, batch=batch, is_training=True, word2ids=word2ids
+        )
+        # summary stats.
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                valid_dl=valid_dl,
+                world_size=world_size,
+                word2ids=word2ids,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer,
+                    "train/valid_",
+                    params.batch_idx_train,
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def get_transducer_model(params: AttributeDict):
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.hidden_dim,
+    )
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        num_layers=params.num_decoder_layers,
+        hidden_dim=params.hidden_dim,
+        embedding_dropout=0.4,
+        rnn_dropout=0.4,
+    )
+    joiner = Joiner(input_dim=params.hidden_dim, output_dim=params.vocab_size)
+    transducer = Transducer(encoder=encoder, decoder=decoder, joiner=joiner)
+
+    return transducer
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+
+    params.update(vars(args))
+    params["env_info"] = get_env_info()
+
+    word2ids = get_word2id(params)
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    # if args.tensorboard and rank == 0:
+    #     tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    # else:
+    #     tb_writer = None
+    tb_writer = None
+
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    else:
+        device = torch.device("cpu")
+    logging.info(f"device: {device}")
+
+    model = get_transducer_model(params)
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    model.device = device
+
+    optimizer = optim.Adam(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    slu = SluDataModule(args)
+    train_dl = slu.train_dataloaders()
+
+    # There are only 60 waves: 30 files are used for training
+    # and the remaining 30 files are used for testing.
+    # We use test data as validation.
+    valid_dl = slu.test_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            word2ids=word2ids,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=None,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    SluDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()

egs/fluent_speech_commands/SLU/transducer/transformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/transformer.py

icefall/shared/make_kn_lm.py

@@ -33,7 +33,7 @@ parser.add_argument(
     "-ngram-order",
     type=int,
     default=4,
-    choices=[2, 3, 4, 5, 6, 7],
+    choices=[1, 2, 3, 4, 5, 6, 7],
     help="Order of n-gram",
 )
 parser.add_argument("-text", type=str, default=None, help="Path to the corpus file")
@@ -105,7 +105,7 @@ class NgramCounts:
     # do as follows: self.counts[3][[5,6,7]][8] += 1.0 where the [3] indexes an
     # array, the [[5,6,7]] indexes a dict, and the [8] indexes a dict.
     def __init__(self, ngram_order, bos_symbol="<s>", eos_symbol="</s>"):
-        assert ngram_order >= 2
+        assert ngram_order >= 1
 
         self.ngram_order = ngram_order
         self.bos_symbol = bos_symbol
@@ -169,7 +169,10 @@ class NgramCounts:
         with open(filename, encoding=default_encoding) as fp:
             for line in fp:
                 line = line.strip(strip_chars)
-                self.add_raw_counts_from_line(line)
+                if self.ngram_order == 1:
+                    self.add_raw_counts_from_line(line.split()[0])
+                else:
+                    self.add_raw_counts_from_line(line)
                 lines_processed += 1
         if lines_processed == 0 or args.verbose > 0:
             print(