update repo

egs/wenetspeech/conformer/local/data.sh

@@ -0,0 +1,110 @@
+#!/usr/bin/env bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+log() {
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%dT%H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+SECONDS=0
+
+# general configuration
+nj=10
+stage=1
+stop_stage=100
+set=L
+data_dir="data"
+
+log "$0 $*"
+. utils/parse_options.sh
+
+. ./path.sh || exit 1;
+. ./cmd.sh || exit 1;
+. ./db.sh || exit 1;
+
+if [ ! -e "${WENETSPEECH}" ]; then
+    log "Fill the value of 'WENETSPEECH' of db.sh"
+    log "or download the data set follwing the instruction in https://wenet-e2e.github.io/WenetSpeech/"
+    exit 1
+fi
+
+if [ ! -d "${WENETSPEECH}/audio" ] && [ ! -f "${WENETSPEECH}/WenetSpeech.json" ]; then
+    echo "Valid WENETSPEECH data not found in ${WENETSPEECH}."
+    echo "Please follow the instruction in https://wenet-e2e.github.io/WenetSpeech/"
+    echo "and re-construct the data."
+    exit 1
+fi
+
+train_set=train_"$(echo "${set}" | tr "[:upper:]" "[:lower:]")"
+dev_set=dev
+test_sets="test_net test_meeting"
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    log "data preparation"
+    mkdir -p ${data_dir}
+    abs_data_dir=$(readlink -f ${data_dir})
+    log "making Kaldi format data directory in ${abs_data_dir}"
+    local/wenetspeech_data_prep.sh \
+        --train-subset ${set} \
+        --stage 1 \
+        ${WENETSPEECH} \
+        ${abs_data_dir}
+
+    # prepare utt2spk and spk2utt files
+    for x in ${train_set} ${dev_set} ${test_sets}; do
+        dir=${data_dir}/${x}
+        paste -d " " <(cut -f 1 ${dir}/segments) <(cut -f 1 ${dir}/segments) | \
+            sort -u > ${dir}/utt2spk
+        utils/utt2spk_to_spk2utt.pl ${dir}/utt2spk > ${dir}/spk2utt
+    done
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    log "process the long term opus audio file, may take about 3 hours"
+    for x in ${train_set} ${dev_set} ${test_sets}; do
+        log "process audio for ${data_dir}/${x}"
+        dir=${data_dir}/${x}
+        mkdir -p ${dir}/logs
+
+        nutt=$(<${dir}/segments wc -l)
+        nj=$((nj<nutt?nj:nutt))
+
+        split_scps=""
+        for n in $(seq ${nj}); do
+            split_scps="${split_scps} ${dir}/logs/segments.${n}"
+        done
+        utils/split_scp.pl ${dir}/segments ${split_scps}
+
+        ${train_cmd} "JOB=1:${nj}" "${dir}/logs/process_audio.JOB.log"\
+            python3 local/process_opus.py \
+                ${dir}/wav.scp \
+                ${dir}/logs/segments.JOB   \
+                ${dir}/logs/wav.JOB.scp
+
+        # modify the `wav.scp` file and rename the `segments` file
+        # rename the `segments` file to avoid the audio file formatting process in stage 3 of `asr.sh`
+        mv ${dir}/wav.scp ${dir}/wav.scp.org
+        mv ${dir}/segments ${dir}/segments.org
+        for n in $(seq ${nj}); do
+            cat ${dir}/logs/wav.${n}.scp || exit 1;
+        done | sort -u > ${dir}/wav.scp
+    done
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    log "format text file"
+    for x in ${train_set} ${dev_set} ${test_sets}; do
+        log "format text for ${data_dir}/${x}"
+        dir=${data_dir}/${x}
+        mv ${dir}/text ${dir}/text.org
+        paste -d " " <(cut -f 1 ${dir}/text.org) \
+            <(cut -f 2- ${dir}/text.org | local/text_normalize.pl) | \
+            sort -u > ${dir}/text
+        utils/fix_data_dir.sh ${dir}
+    done
+fi
+
+log "Successfully finished. [elapsed=${SECONDS}s]"

egs/wenetspeech/conformer/local/extract_meta.py

@@ -13,20 +13,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import sys
-import os
 import argparse
 import json
+import os
+import sys
 
 
 def get_args():
-    parser = argparse.ArgumentParser(description="""
+    parser = argparse.ArgumentParser(
+        description="""
       This script is used to process raw json dataset of WenetSpeech,
       where the long wav is splitinto segments and
       data of wenet format is generated.
-      """)
-    parser.add_argument('input_json', help="""Input json file of WenetSpeech""")
-    parser.add_argument('output_dir', help="""Output dir for prepared data""")
+      """
+    )
+    parser.add_argument("input_json", help="""Input json file of WenetSpeech""")
+    parser.add_argument("output_dir", help="""Output dir for prepared data""")
 
     args = parser.parse_args()
     return args
@@ -39,58 +41,68 @@ def meta_analysis(input_json, output_dir):
         os.makedirs(output_dir)
 
     try:
-        with open(input_json, 'r') as injson:
+        with open(input_json, "r") as injson:
             json_data = json.load(injson)
     except Exception:
-        sys.exit(f'Failed to load input json file: {input_json}')
+        sys.exit(f"Failed to load input json file: {input_json}")
     else:
-        if json_data['audios'] is not None:
-            with open(f'{output_dir}/text', 'w') as utt2text, \
-                 open(f'{output_dir}/segments', 'w') as segments, \
-                 open(f'{output_dir}/utt2dur', 'w') as utt2dur, \
-                 open(f'{output_dir}/wav.scp', 'w') as wavscp, \
-                 open(f'{output_dir}/utt2subsets', 'w') as utt2subsets, \
-                 open(f'{output_dir}/reco2dur', 'w') as reco2dur:
-                for long_audio in json_data['audios']:
+        if json_data["audios"] is not None:
+            with open(f"{output_dir}/text", "w") as utt2text, open(
+                f"{output_dir}/segments", "w"
+            ) as segments, open(f"{output_dir}/utt2dur", "w") as utt2dur, open(
+                f"{output_dir}/wav.scp", "w"
+            ) as wavscp, open(
+                f"{output_dir}/utt2subsets", "w"
+            ) as utt2subsets, open(
+                f"{output_dir}/reco2dur", "w"
+            ) as reco2dur:
+                for long_audio in json_data["audios"]:
                     try:
                         long_audio_path = os.path.realpath(
-                            os.path.join(input_dir, long_audio['path']))
-                        aid = long_audio['aid']
-                        segments_lists = long_audio['segments']
-                        duration = long_audio['duration']
-                        assert (os.path.exists(long_audio_path))
+                            os.path.join(input_dir, long_audio["path"])
+                        )
+                        aid = long_audio["aid"]
+                        segments_lists = long_audio["segments"]
+                        duration = long_audio["duration"]
+                        assert os.path.exists(long_audio_path)
                     except AssertionError:
-                        print(f'''Warning: {aid} something is wrong,
-                                  maybe AssertionError, skipped''')
+                        print(
+                            f"""Warning: {aid} something is wrong,
+                                  maybe AssertionError, skipped"""
+                        )
                         continue
                     except Exception:
-                        print(f'''Warning: {aid} something is wrong, maybe the
-                                  error path: {long_audio_path}, skipped''')
+                        print(
+                            f"""Warning: {aid} something is wrong, maybe the
+                                  error path: {long_audio_path}, skipped"""
+                        )
                         continue
                     else:
-                        wavscp.write(f'{aid}\t{long_audio_path}\n')
-                        reco2dur.write(f'{aid}\t{duration}\n')
+                        wavscp.write(f"{aid}\t{long_audio_path}\n")
+                        reco2dur.write(f"{aid}\t{duration}\n")
                         for segment_file in segments_lists:
                             try:
-                                sid = segment_file['sid']
-                                start_time = segment_file['begin_time']
-                                end_time = segment_file['end_time']
+                                sid = segment_file["sid"]
+                                start_time = segment_file["begin_time"]
+                                end_time = segment_file["end_time"]
                                 dur = end_time - start_time
-                                text = segment_file['text']
+                                text = segment_file["text"]
                                 segment_subsets = segment_file["subsets"]
                             except Exception:
-                                print(f'''Warning: {segment_file} something
-                                          is wrong, skipped''')
+                                print(
+                                    f"""Warning: {segment_file} something
+                                          is wrong, skipped"""
+                                )
                                 continue
                             else:
-                                utt2text.write(f'{sid}\t{text}\n')
+                                utt2text.write(f"{sid}\t{text}\n")
                                 segments.write(
-                                    f'{sid}\t{aid}\t{start_time}\t{end_time}\n'
+                                    f"{sid}\t{aid}\t{start_time}\t{end_time}\n"
                                 )
-                                utt2dur.write(f'{sid}\t{dur}\n')
+                                utt2dur.write(f"{sid}\t{dur}\n")
                                 segment_sub_names = " ".join(segment_subsets)
-                                utt2subsets.write(
-                                    f'{sid}\t{segment_sub_names}\n')
+                                utt2subsets.write(f"{sid}\t{segment_sub_names}\n")
+
 
 def main():
     args = get_args()
@@ -98,5 +110,5 @@ def main():
     meta_analysis(args.input_json, args.output_dir)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

egs/wenetspeech/conformer/local/process_opus.py

@@ -16,14 +16,15 @@
 
 # usage: python3 process_opus.py wav.scp segments output_wav.scp
 
-from pydub import AudioSegment
-import sys
 import os
+import sys
+
+from pydub import AudioSegment
 
 
 def read_file(wav_scp, segments):
     wav_scp_dict = {}
-    with open(wav_scp, 'r', encoding='UTF-8') as fin:
+    with open(wav_scp, "r", encoding="UTF-8") as fin:
         for line_str in fin:
             wav_id, path = line_str.strip().split()
             wav_scp_dict[wav_id] = path
@@ -32,7 +33,7 @@ def read_file(wav_scp, segments):
     seg_path_list = []
     start_time_list = []
     end_time_list = []
-    with open(segments, 'r', encoding='UTF-8') as fin:
+    with open(segments, "r", encoding="UTF-8") as fin:
         for line_str in fin:
             arr = line_str.strip().split()
             assert len(arr) == 4
@@ -44,30 +45,27 @@ def read_file(wav_scp, segments):
 
 
 # TODO(Qijie): Fix the process logic
-def output(output_wav_scp, utt_list, seg_path_list, start_time_list,
-           end_time_list):
+def output(output_wav_scp, utt_list, seg_path_list, start_time_list, end_time_list):
     num_utts = len(utt_list)
     step = int(num_utts * 0.01)
-    with open(output_wav_scp, 'w', encoding='UTF-8') as fout:
+    with open(output_wav_scp, "w", encoding="UTF-8") as fout:
         previous_wav_path = ""
         for i in range(num_utts):
             utt_id = utt_list[i]
             current_wav_path = seg_path_list[i]
-            output_dir = (os.path.dirname(current_wav_path)) \
-                .replace("audio", 'audio_seg')
-            seg_wav_path = os.path.join(output_dir, utt_id + '.wav')
-
-            # if not os.path.exists(output_dir):
-            #     os.makedirs(output_dir)
+            output_dir = (os.path.dirname(current_wav_path)).replace(
+                "audio", "audio_seg"
+            )
+            seg_wav_path = os.path.join(output_dir, utt_id + ".wav")
 
+            os.makedirs(output_dir, exist_ok=True)
             if current_wav_path != previous_wav_path:
                 source_wav = AudioSegment.from_file(current_wav_path)
             previous_wav_path = current_wav_path
 
             start = int(start_time_list[i] * 1000)
             end = int(end_time_list[i] * 1000)
-            target_audio = source_wav[start:end].set_frame_rate(16000) \
-                .set_sample_width(2)
+            target_audio = source_wav[start:end].set_frame_rate(16000)
             target_audio.export(seg_wav_path, format="wav")
 
             fout.write("{} {}\n".format(utt_id, seg_wav_path))
@@ -80,11 +78,11 @@ def main():
     segments = sys.argv[2]
     output_wav_scp = sys.argv[3]
 
-    utt_list, seg_path_list, start_time_list, end_time_list \
-        = read_file(wav_scp, segments)
-    output(output_wav_scp, utt_list, seg_path_list, start_time_list,
-           end_time_list)
+    utt_list, seg_path_list, start_time_list, end_time_list = read_file(
+        wav_scp, segments
+    )
+    output(output_wav_scp, utt_list, seg_path_list, start_time_list, end_time_list)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

egs/wenetspeech/conformer/local/text_normalize.pl

@@ -0,0 +1,24 @@
+#!/usr/bin/env perl
+use utf8;
+use open qw(:std :utf8);
+use warnings;
+
+while (<STDIN>) {
+    chomp;
+    # remove non UTF-8 whitespace character
+    if ($_ =~ /　/) {$_ =~ s:　::g;}
+    if ($_ =~ / /) {$_ =~ s: ::g;}
+    # upper letters
+    if ($_ =~ /[a-zA-Z]/) {$_ =~ uc $_;}
+    # add "_" before and after each English word
+    if ($_ =~ /([A-Z]+)\s+([A-Z]+)/) {$_ =~ s/([A-Z]+)\s+([A-Z]+)/$1\_$2/g;}
+    if ($_ =~ /([A-Z]+)\s+([A-Z]+)/) {$_ =~ s/([A-Z]+)\s+([A-Z]+)/$1\_$2/g;}
+    if ($_ =~ m/([A-Z]+)(\p{Han}+)/) {$_ =~ s/([A-Z]+)(\p{Han}+)/$1\_$2/g;}
+    if ($_ =~ m/(\p{Han}+)([A-Z]+)/) {$_ =~ s/(\p{Han}+)([A-Z]+)/$1\_$2/g;}
+    # remove UTF-8 whitespace charcter
+    if ($_ =~ /\s+/) {$_ =~ s:\s+::g;}
+    # replace "_" with a normal whitespace
+    if ($_ =~ /\_/) {$_ =~ s:\_: :g;}
+
+    print "$_\n";
+}

egs/wenetspeech/conformer/local/wenetspeech_data_prep.sh

@@ -24,7 +24,7 @@ stage=1
 prefix=
 train_subset=L
 
-. ./utils/parse_options.sh || exit 1;
+. utils/parse_options.sh || exit 1;
 
 filter_by_id () {
   idlist=$1

egs/wenetspeech/conformer/run.sh

@@ -41,6 +41,7 @@ set -e
 set -u
 set -o pipefail
 
+set=L
 train_set=train_l
 valid_set=dev
 test_sets="dev test_net test_meeting"
@@ -71,15 +72,15 @@ fi
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     echo "stage 0: Data preparation"
     # Data preparation
-    local/wenetspeech_data_prep.sh $raw_data $feats_dir
-    mkdir $feats_dir/data
-    mv $feats_dir/$train_set $feats_dir/data/$train_set
-    for x in $test_sets; do
-        mv $feats_dir/$x $feats_dir/data/
-    done
+    local/data.sh "--set ${set}"
+#    mkdir $feats_dir/data
+#    mv $feats_dir/$train_set $feats_dir/data/$train_set
+#    for x in $test_sets; do
+#        mv $feats_dir/$x $feats_dir/data/
+#    done
 fi
 
-if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    echo "stage 1: Feature and CMVN Generation"
-    utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 0.1
-fi
+#if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+#    echo "stage 1: Feature and CMVN Generation"
+#    utils/compute_cmvn.sh --fbankdir ${feats_dir}/data/${train_set} --cmd "$train_cmd" --nj $nj --feats_dim ${feats_dim} --config_file "$asr_config" --scale 0.1
+#fi