updated

egs/swbd/ASR/conformer_ctc/asr_datamodule.py

@@ -225,6 +225,8 @@ class SwitchBoardAsrDataModule:
         else:
             logging.info("Disable MUSAN")
 
+        cuts_train = cuts_train.trim_to_supervisions(keep_overlapping=False)
+
         if self.args.concatenate_cuts:
             logging.info(
                 f"Using cut concatenation with duration factor "
@@ -392,25 +394,23 @@ class SwitchBoardAsrDataModule:
     def train_all_cuts(self) -> CutSet:
         logging.info("switchboard: About to get train cuts")
         return (
-            load_manifest_lazy(self.args.manifest_dir / "swbd_cuts_all.jsonl.gz")
+            load_manifest_lazy(self.args.manifest_dir / "swbd" / "swbd_cuts_all.jsonl.gz")
             .subset(last=2388)
-            .trim_to_supervisions(keep_all_channels=True)
         )
 
     @lru_cache()
     def dev_cuts(self) -> CutSet:
         logging.info("switchboard: About to get dev cuts")
         return (
-            load_manifest_lazy(self.args.manifest_dir / "swbd_cuts_all.jsonl.gz")
+            load_manifest_lazy(self.args.manifest_dir / "swbd" / "swbd_cuts_all.jsonl.gz")
             .subset(first=50)
-            .trim_to_supervisions(keep_all_channels=True)
         )
 
     @lru_cache()
     def test_eval2000_cuts(self) -> CutSet:
         logging.info("switchboard: About to get eval2000 cuts")
         return load_manifest_lazy(
-            self.args.manifest_dir / "swbd_cuts_eval2000.jsonl.gz"
+            self.args.manifest_dir / "eval2000" / "eval2000_cuts_all.jsonl.gz"
         )
 
     @lru_cache()

egs/swbd/ASR/local/compute_fbank_swbd.py

@@ -66,7 +66,7 @@ def get_args():
     parser.add_argument(
         "--perturb-speed",
         type=str2bool,
-        default=True,
+        default=False,
         help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
     )
 
@@ -90,7 +90,7 @@ def compute_fbank_switchboard(
         sp.load(bpe_model)
 
     if dataset is None:
-        dataset_parts = ("all")
+        dataset_parts = ("all",)
     else:
         dataset_parts = dataset.split(" ", -1)
 
@@ -152,7 +152,7 @@ if __name__ == "__main__":
     logging.basicConfig(format=formatter, level=logging.INFO)
     args = get_args()
     logging.info(vars(args))
-    for dir_name in ["swbd"]:
+    for dir_name in ["swbd", "eval2000"]:
         compute_fbank_switchboard(
             dir_name=dir_name,
             bpe_model=args.bpe_model,

egs/swbd/ASR/local/normalize_and_filter_supervisions.py

@@ -119,6 +119,9 @@ class FisherSwbdNormalizer:
             (re.compile(r"(\[.*?\])-"), r"\1"),
             # Just remove all dashes
             (re.compile(r"-"), r" "),
+
+            # Fix an issue related to [vocalized-noise]
+            (re.compile(r"\[vocalized noise\]"), r"\[vocalized-noise\]"),
         ]
 
         # unwanted symbols in the transcripts
@@ -153,7 +156,7 @@ class FisherSwbdNormalizer:
         # then clean up whitespace
         text = self.whitespace_regexp.sub(" ", text).strip()
 
-        return text
+        return text.upper()
 # fmt: on
 
 
@@ -189,6 +192,7 @@ def main():
                 continue
 
             writer.write(sup)
+    print(f"tot: {tot}, skip: {skip}")
 
 
 def test():
@@ -224,5 +228,5 @@ def test():
 
 
 if __name__ == "__main__":
-    test()
-#    main()
+    # test(); exit()
+    main()

egs/swbd/ASR/local/prepare_lang.py

@@ -249,7 +249,7 @@ def lexicon_to_fst(
     lexicon: Lexicon,
     token2id: Dict[str, int],
     word2id: Dict[str, int],
-    sil_token: str = "sil",
+    sil_token: str = "SIL",
     sil_prob: float = 0.5,
     need_self_loops: bool = False,
 ) -> k2.Fsa:
@@ -346,7 +346,7 @@ def main():
     args = get_args()
     lang_dir = Path(args.lang_dir)
     lexicon_filename = lang_dir / "lexicon.txt"
-    sil_token = "sil"
+    sil_token = "SIL"
     sil_prob = 0.5
 
     lexicon = read_lexicon(lexicon_filename)

egs/swbd/ASR/local/prepare_lang_bpe.py

@@ -178,7 +178,7 @@ def get_args():
     parser.add_argument(
         "--oov",
         type=str,
-        default="<unk>",
+        default="<UNK>",
         help="The out of vocabulary word in lexicon.",
     )
 

egs/swbd/ASR/prepare.sh

@@ -23,7 +23,8 @@ stop_stage=100
 #     - speech
 
 dl_dir=./download
-swbd1_dir="/export/corpora3/LDC/LDC97S62"
+# swbd1_dir="/export/corpora3/LDC/LDC97S62"
+swbd1_dir=./download/LDC97S62/
 
 # eval2000_dir contains the following files and directories
 # downloaded from LDC website:
@@ -70,15 +71,14 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
     if [ ! -e data/manifests/.swbd.done ]; then
         lhotse prepare switchboard --absolute-paths 1 --omit-silence $swbd1_dir data/manifests/swbd
         ./local/normalize_and_filter_supervisions.py \
-            data/manifests/swbd/swbd_supervisions.jsonl \
-            data/manifests/swbd/swbd_supervisions_norm.jsonl
-        cp data/manifests/swbd/swbd_recordings.jsonl data/manifests/recordings_swbd.jsonl
+            data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+            data/manifests/swbd/swbd_supervisions_all_norm.jsonl.gz
+        mv data/manifests/swbd/swbd_supervisions_all_norm.jsonl.gz data/manifests/swbd/swbd_supervisions_all.jsonl.gz
 
-
-        lhotse prepare $eval2000_dir data/manifests_eval2000
+        lhotse prepare eval2000 --absolute-paths 1 $eval2000_dir data/manifests/eval2000
         ./local/normalize_eval2000.py \
-            data/manifests_eval2000/eval2000_supervisions_unnorm.jsonl.gz \
-            data/manifests_eval2000/eval2000_supervisions_norm.jsonl.gz
+            data/manifests/eval2000/eval2000_supervisions_unnorm.jsonl.gz \
+            data/manifests/eval2000/eval2000_supervisions_all.jsonl.gz
 
         ./local/rt03_data_prep.sh $rt03_dir
 
@@ -96,20 +96,6 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
             rm data/local/${x}/text.org
         done
 
-        python ./local/filter_empty_text.py --kaldi-data-dir data/local/eval2000
-        ./utils/fix_data_dir.sh data/local/eval2000
-        lhotse kaldi import data/local/eval2000 8000 data/manifests_eval2000
-        mv data/manifests_eval2000/recordings.jsonl.gz data/manifests_eval2000/swbd_recordings_eval2000.jsonl.gz
-        mv data/manifests_eval2000/supervisions.jsonl.gz data/manifests_eval2000/swbd_supervisions_eval2000.jsonl.gz
-
-        python ./local/filter_empty_text.py --kaldi-data-dir data/local/rt03
-        ./utils/fix_data_dir.sh data/local/rt03
-        lhotse kaldi import data/local/rt03 8000 data/manifests_rt03
-        mv data/manifests_rt03/recordings.jsonl.gz data/manifests_rt03/swbd_recordings_rt03.jsonl.gz
-        mv data/manifests_rt03/supervisions.jsonl.gz data/manifests_rt03/swbd_supervisions_rt03.jsonl.gz
-
-        lhotse fix data/manifests_train/swbd_recordings_all.jsonl.gz data/manifests_train/swbd_supervisions_all.jsonl.gz data/manifests
-        lhotse fix data/manifests_eval2000/swbd_recordings_eval2000.jsonl.gz data/manifests_eval2000/swbd_supervisions_eval2000.jsonl.gz data/manifests
         lhotse fix data/manifests_rt03/swbd_recordings_rt03.jsonl.gz data/manifests_rt03/swbd_supervisions_rt03.jsonl.gz data/manifests
 
         touch data/manifests/.swbd.done
@@ -128,7 +114,7 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
 fi
 
 if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
-    log "Stage 3: Compute fbank for switchboard"
+    log "Stage 3: Compute fbank for SwitchBoard"
     mkdir -p data/fbank
     if [ ! -e data/fbank/.swbd.done ]; then
         ./local/compute_fbank_swbd.py
@@ -150,13 +136,29 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     lang_dir=data/lang_phone
     mkdir -p $lang_dir
 
+    if ! which jq; then
+      echo "This script is intended to be used with jq but you have not installed jq
+      Note: in Linux, you can install jq with the following command:
+      1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+      2. chmod +x ./jq
+      3. cp jq /usr/bin" && exit 1
+    fi
+    if [ ! -f $lang_dir/text ] || [ ! -s $lang_dir/text ]; then
+        log "Prepare text."
+        gunzip -c data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+        | jq '.text' | sed 's/"//g'  > $lang_dir/text
+    fi
+
     log "prepare dict"
-    cut -f 2- -d" " data/local/train/text >${lang_dir}/input.txt
+    ./local/swbd1_prepare_dict.sh $swbd1_dir
+    cut -f 2- -d" " $lang_dir/text >${lang_dir}/input.txt
     # [noise] nsn
     # !sil sil
     # <unk> spn
     cat data/local/dict_nosp/lexicon.txt |
-        sort | uniq >$lang_dir/lexicon.txt
+        sort | uniq >$lang_dir/lexicon_lower.txt
+
+    cat $lang_dir/lexicon_lower.txt | tr a-z A-Z > $lang_dir/lexicon.txt
 
     if [ ! -f $lang_dir/L_disambig.pt ]; then
         ./local/prepare_lang.py --lang-dir $lang_dir
@@ -192,7 +194,7 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
         if [ ! -f $lang_dir/transcript_words.txt ]; then
             log "Generate data for BPE training"
 
-            cat ./data/local/train/text | cut -d " " -f 2- >$lang_dir/transcript_words.txt
+            cat data/lang_phone/text | cut -d " " -f 2- >$lang_dir/transcript_words.txt
         fi
 
         if [ ! -f $lang_dir/bpe.model ]; then
@@ -239,7 +241,7 @@ if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
             ./local/convert_transcript_words_to_tokens.py \
                 --lexicon $lang_dir/lexicon.txt \
                 --transcript $lang_dir/transcript_words.txt \
-                --oov "<unk>" \
+                --oov "<UNK>" \
                 >$lang_dir/transcript_tokens.txt
         fi