data process for ted-en-zh st1

examples/ted_en_zh/st1/local/data.sh

@@ -2,16 +2,18 @@
 
 set -e
 
-stage=-1
+stage=1
 stop_stage=100
 dict_dir=data/lang_char
 
 # bpemode (unigram or bpe)
 nbpe=8000
-bpemode=unigram
+bpemode=bpe
 bpeprefix="${dict_dir}/bpe_${bpemode}_${nbpe}"
 data_dir=./TED_EnZh
-
+target_dir=data/ted_en_zh
+dumpdir=data/dump
+do_delta=false
 
 source ${MAIN_ROOT}/utils/parse_options.sh
 
@@ -38,75 +40,163 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
         exit 1
     fi
 
-    # generate manifests
-    python3 ${TARGET_DIR}/ted_en_zh/ted_en_zh.py \
-    --manifest_prefix="data/manifest" \
-    --src_dir="${data_dir}"
+    # # extract data 
+    # echo "data Extraction"
+    # python3 local/ted_en_zh.py \
+    # --tgt-dir=${target_dir} \
+    # --src-dir=${data_dir}
 
-    echo "Complete raw data pre-process."
 fi
-
+prep_dir=${target_dir}/data_prep 
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
-    # compute mean and stddev for normalizer
-    num_workers=$(nproc)
-    python3 ${MAIN_ROOT}/utils/compute_mean_std.py \
-    --manifest_path="data/manifest.train.raw" \
-    --num_samples=-1 \
-    --spectrum_type="fbank" \
-    --feat_dim=80 \
-    --delta_delta=false \
-    --sample_rate=16000 \
-    --stride_ms=10.0 \
-    --window_ms=25.0 \
-    --use_dB_normalization=False \
-    --num_workers=${num_workers} \
-    --output_path="data/mean_std.json"
-
-    if [ $? -ne 0 ]; then
-        echo "Compute mean and stddev failed. Terminated."
-        exit 1
-    fi
+    ### Task dependent. You have to make data the following preparation part by yourself.
+    ### But you can utilize Kaldi recipes in most cases
+    echo "stage 0: Data preparation"
+    for set in train dev test; do
+    # for set in train; do
+        dst=${target_dir}/${set}
+        for lang in en zh; do
+
+            if [ ${lang} = 'en' ]; then
+                echo "remove punctuation $lang"
+                # remove punctuation
+                local/remove_punctuation.pl < ${dst}/${lang}.org > ${dst}/${lang}.raw
+            else
+                cp ${dst}/${lang}.org ${dst}/${lang}.raw
+            fi
+
+            paste -d " " ${dst}/.yaml ${dst}/${lang}.raw | sort > ${dst}/text.${lang}
+
+
+        done
+        # error check
+        n=$(cat ${dst}/.yaml | wc -l)
+        n_en=$(cat ${dst}/en.raw | wc -l)
+        n_tgt=$(cat ${dst}/zh.raw | wc -l)
+        [ ${n} -ne ${n_en} ] && echo "Warning: expected ${n} data data files, found ${n_en}" && exit 1;
+        [ ${n} -ne ${n_tgt} ] && echo "Warning: expected ${n} data data files, found ${n_tgt}" && exit 1;
+
+        echo "done text processing"
+        cat ${dst}/wav.scp.org | uniq | sort -k1,1 -u > ${dst}/wav.scp
+        cat ${dst}/utt2spk.org | uniq | sort -k1,1 -u > ${dst}/utt2spk
+
+        cat ${dst}/utt2spk | utils/utt2spk_to_spk2utt.pl | sort -k1,1 -u > ${dst}/spk2utt
+        rm -rf ${prep_dir}/${set}.en-zh
+        mkdir -p ${prep_dir}/${set}.en-zh
+        echo "remove duplicate lines..."
+        cut -d ' ' -f 1 ${dst}/text.en | sort | uniq -c | sort -n -k1 -r | grep -v '1 ted-en-zh' \
+            | sed 's/^[ \t]*//' > ${dst}/duplicate_lines
+        cut -d ' ' -f 1 ${dst}/text.en | sort | uniq -c | sort -n -k1 -r | grep '1 ted-en-zh' \
+            | cut -d '1' -f 2- | sed 's/^[ \t]*//' > ${dst}/reclist
+        reduce_data_dir.sh ${dst} ${dst}/reclist ${prep_dir}/${set}.en-zh
+        echo "done wav processing"
+        for l in en zh; do
+            cp ${dst}/text.${l} ${prep_dir}/${set}.en-zh/text.${l}
+        done
+        utils/fix_data_dir.sh --utt_extra_files \
+        "text.en text.zh" \
+        ${prep_dir}/${set}.en-zh
+    done
 fi
 
+feat_tr_dir=${dumpdir}/train/delta${do_delta}; mkdir -p ${feat_tr_dir}
+feat_dt_dir=${dumpdir}/dev/delta${do_delta}; mkdir -p ${feat_dt_dir}
+feat_trans_dir=${dumpdir}/test/delta${do_delta}; mkdir -p ${feat_trans_dir}
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # build vocabulary
-    python3 ${MAIN_ROOT}/utils/build_vocab.py \
-    --unit_type "spm" \
-    --spm_vocab_size=${nbpe} \
-    --spm_mode ${bpemode} \
-    --spm_model_prefix ${bpeprefix} \
-    --vocab_path="${dict_dir}/vocab.txt" \
-    --text_keys 'text' 'text1' \
-    --manifest_paths="data/manifest.train.raw"
-
-
-    if [ $? -ne 0 ]; then
-        echo "Build vocabulary failed. Terminated."
-        exit 1
-    fi
+    ### Task dependent. You have to design training and dev sets by yourself.
+    ### But you can utilize Kaldi recipes in most cases
+    echo "stage 1: Feature Generation"
+    fbankdir=data/fbank
+    # Generate the fbank features; by default 80-dimensional fbanks with pitch on each frame
+    for x in train dev test; do
+        steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 32 --write_utt2num_frames true \
+            ${prep_dir}/${x}.en-zh data/make_fbank/${x} ${fbankdir}
+    done
+    
+    echo "speed perturbation"
+    utils/perturb_data_dir_speed.sh 0.9 ${prep_dir}/train.en-zh ${prep_dir}/temp1.en-zh
+    utils/perturb_data_dir_speed.sh 1.0 ${prep_dir}/train.en-zh ${prep_dir}/temp2.en-zh
+    utils/perturb_data_dir_speed.sh 1.1 ${prep_dir}/train.en-zh ${prep_dir}/temp3.en-zh
+
+    utils/combine_data.sh --extra-files utt2uniq ${prep_dir}/train_sp.en-zh \
+    ${prep_dir}/temp1.en-zh ${prep_dir}/temp2.en-zh ${prep_dir}/temp3.en-zh
+    rm -r ${prep_dir}/temp*.en-zh 
+    utils/fix_data_dir.sh ${prep_dir}/train_sp.en-zh
+
+    steps/make_fbank_pitch.sh --cmd "$train_cmd" --nj 32 --write_utt2num_frames true \
+        ${prep_dir}/train_sp.en-zh exp/make_fbank/train_sp.en-zh ${fbankdir}
+
+    for lang in en zh; do
+        cat /dev/null > ${prep_dir}/train_sp.en-zh/text.${lang}
+        for p in "sp0.9-" "sp1.0-" "sp1.1-"; do
+            awk -v p=${p} '{printf("%s %s%s\n", $1, p, $1);}' ${prep_dir}/train.en-zh/utt2spk > ${prep_dir}/train_sp.en-zh/utt_map
+            utils/apply_map.pl -f 1 ${prep_dir}/train_sp.en-zh/utt_map < ${prep_dir}/train.en-zh/text.${lang} >>${prep_dir}/train_sp.en-zh/text.${lang}
+        done
+    done
+
+    for x in train_sp dev test; do
+        local/divide_lang.sh ${prep_dir}/${x}.en-zh zh
+    done
+
+    for x in train_sp dev; do
+        # remove utt having more than 3000 frames
+        # remove utt having more than 400 characters
+        for lang in zh en; do
+            remove_longshortdata.sh --maxframes 3000 --maxchars 400 ${prep_dir}/${x}.en-zh.${lang} ${prep_dir}/${x}.en-zh.${lang}.tmp
+        done
+        cut -f 1 -d " " ${prep_dir}/${x}.en-zh.en.tmp/text > ${prep_dir}/${x}.en-zh.${lang}.tmp/reclist1
+        cut -f 1 -d " " ${prep_dir}/${x}.en-zh.${lang}.tmp/text > ${prep_dir}/${x}.en-zh.${lang}.tmp/reclist2
+        comm -12 ${prep_dir}/${x}.en-zh.${lang}.tmp/reclist1 ${prep_dir}/${x}.en-zh.${lang}.tmp/reclist2 > ${prep_dir}/${x}.en-zh.en.tmp/reclist
+
+        for lang in zh en; do
+            reduce_data_dir.sh ${prep_dir}/${x}.en-zh.${lang}.tmp ${prep_dir}/${x}.en-zh.en.tmp/reclist ${prep_dir}/${x}.en-zh.${lang}
+            utils/fix_data_dir.sh  ${prep_dir}/${x}.en-zh.${lang}
+        done
+        rm -rf ${prep_dir}/${x}.en-zh.*.tmp
+    done
+
+    compute-cmvn-stats scp:${prep_dir}/train_sp.en-zh/feats.scp ${prep_dir}/train_sp.en-zh/cmvn.ark
+
+    dump.sh --cmd "$train_cmd" --nj 80 --do_delta $do_delta \
+        ${prep_dir}/train_sp.en-zh/feats.scp ${prep_dir}/train_sp.en-zh/cmvn.ark ${prep_dir}/dump_feats/train_sp.en-zh ${feat_tr_dir}
+    dump.sh --cmd "$train_cmd" --nj 32 --do_delta $do_delta \
+        ${prep_dir}/dev.en-zh/feats.scp ${prep_dir}/train_sp.en-zh/cmvn.ark ${prep_dir}/dump_feats/dev.en-zh ${feat_dt_dir}
+    dump.sh --cmd "$train_cmd" --nj 32 --do_delta $do_delta \
+        ${prep_dir}/test.en-zh/feats.scp ${prep_dir}/train_sp.en-zh/cmvn.ark ${prep_dir}/dump_feats/test.en-zh ${feat_trans_dir}
 fi
 
+dict=${dict_dir}/ted_en_zh_${bpemode}${nbpe}_joint.txt
+nlsyms=${dict_dir}/ted_en_zh_non_lang_syms.txt
+bpemodel=${dict_dir}/ted_en_zh_${bpemode}${nbpe}
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
-    # format manifest with tokenids, vocab size
-    for set in train dev test; do
-    {
-        python3 ${MAIN_ROOT}/utils/format_triplet_data.py \
-        --feat_type "raw" \
-        --cmvn_path "data/mean_std.json" \
-        --unit_type "spm" \
-        --spm_model_prefix ${bpeprefix} \
-        --vocab_path="${dict_dir}/vocab.txt" \
-        --manifest_path="data/manifest.${set}.raw" \
-        --output_path="data/manifest.${set}"
-
-        if [ $? -ne 0 ]; then
-            echo "Formt mnaifest failed. Terminated."
-            exit 1
-        fi
-    }&
+    echo "stage 2: Dictionary and Json Data Preparation"
+    # echo "make a non-linguistic symbol list for all languages"
+    # grep sp1.0 ${prep_dir}/train_sp.en-zh.*/text | cut -f 2- -d' ' | grep -o -P '&[^;];'| sort | uniq > ${nlsyms}
+    # cat ${nlsyms}
+
+    echo "make a joint source and target dictionary"
+    echo "<unk> 1" > ${dict} # <unk> must be 1, 0 will be used for "blank" in CTC
+    offset=$(wc -l < ${dict})
+    grep sp1.0 ${prep_dir}/train_sp.en-zh.*/text | cut -f 2- -d' ' | grep -v -e '^\s*$' > ${dict_dir}/input.txt
+    spm_train  --input=${dict_dir}/input.txt --vocab_size=${nbpe} --model_type=${bpemode} --model_prefix=${bpemodel} --input_sentence_size=100000000 --character_coverage=1.0
+    spm_encode --model=${bpemodel}.model --output_format=piece < ${dict_dir}/input.txt | tr ' ' '\n' | sort | uniq | awk -v offset=${offset} '{print $0 " " NR+offset}' >> ${dict}
+    wc -l ${dict}
+
+    echo "make json files"
+    data2json.sh --nj 16 --feat ${feat_tr_dir}/feats.scp --text ${prep_dir}/train_sp.en-zh.zh/text --bpecode ${bpemodel}.model --lang zh \
+        ${prep_dir}/train_sp.en-zh.zh ${dict} > ${feat_tr_dir}/data_${bpemode}${nbpe}.json
+    data2json.sh --feat ${feat_dt_dir}/feats.scp --text ${prep_dir}/dev.en-zh.zh/text --bpecode ${bpemodel}.model --lang zh \
+        ${prep_dir}/dev.en-zh.zh ${dict} > ${feat_dt_dir}/data_${bpemode}${nbpe}.json
+    data2json.sh --feat ${feat_dt_dir}/feats.scp --text ${prep_dir}/test.en-zh.zh/text --bpecode ${bpemodel}.model --lang zh \
+        ${prep_dir}/test.en-zh.zh ${dict} > ${feat_trans_dir}/data_${bpemode}${nbpe}.json
+    echo "update json (add source references)"
+    # update json (add source references)
+    for x in ${train_set} ${train_dev}; do
+        feat_dir=${dumpdir}/${x}/delta${do_delta}
+        data_dir=data/$(echo ${x} | cut -f 1 -d ".").en-zh.en
+        update_json.sh --text ${data_dir}/text.${src_case} --bpecode ${bpemodel}.model \
+            ${feat_dir}/data_${bpemode}${nbpe}.json ${data_dir} ${dict}
     done
-    wait
 fi
-
 echo "Ted En-Zh Data preparation done."
 exit 0

examples/ted_en_zh/st1/path.sh

 export MAIN_ROOT=`realpath ${PWD}/../../../`
 
-export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
+export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PWD}/utils:${PATH}
 export LC_ALL=C
 
 export PYTHONDONTWRITEBYTECODE=1
@@ -13,3 +13,11 @@ export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/lib/
 
 MODEL=u2_st
 export BIN_DIR=${MAIN_ROOT}/paddlespeech/s2t/exps/${MODEL}/bin
+
+# Kaldi
+export KALDI_ROOT=${MAIN_ROOT}/tools/kaldi
+[ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh
+export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$PWD:$PATH
+[ ! -f $KALDI_ROOT/tools/config/common_path.sh ] && echo >&2 "The standard file $KALDI_ROOT/tools/config/common_path.sh is not present, can not using Kaldi!"
+[ -f $KALDI_ROOT/tools/config/common_path.sh ] && . $KALDI_ROOT/tools/config/common_path.sh
+export train_cmd="run.pl"
\ No newline at end of file