きりたん歌唱データベースにjuliusの音素アライメントを適用させるコードメモ

01.py

import subprocess
from pathlib import Path

lab = Path("01.lab.txt").read_text()

i = 0
out_labs = []
for l in lab.splitlines():
    a, b, c = l.split()
    if c == "pau":
        if len(out_labs) == 0:
            continue

        start = float(out_labs[0].split()[0])
        end = float(b)
        subprocess.run(
            f"sox w1.wav split-{i}.wav trim {start - 0.1} {end - start}", shell=True
        )

        for ii in range(len(out_labs)):
            s, e, p = out_labs[ii].split()
            out_labs[ii] = f"0 0 {p}"

        Path(f"split-{i}.lab").write_text("\n".join(out_labs))
        out_labs = []
        i += 1

    else:
        out_labs.append(l)

02.bash

sox 01.wav -r 16k -c 1 -b 16 w1.wav

for i in {0..8}; do

cat split-$i.lab |\
awk '{print $NF}' |\
sed -r 's/pau/sp/g' |\
sed -r 's/br/sp/g' |\
sed -r 's/cl/q/g' |\
tr '\n' ' ' \
> /tmp/yomi.txt

julius4seg_segment \
    split-$i.wav \
    /tmp/yomi.txt \
    output-split-$i.lab \
    --input_yomi_type phoneme \
    --hmm_model dictation-kit/model/dnn/binhmm.SID \
    --only_2nd_path \
    --model_type dnn \
    --options='
        -hlist dictation-kit/model/dnn/logicalTri.bin
        -dnnconf dictation-kit/julius.dnnconf
    '

done

03.py

from pathlib import Path

base_lab = Path("01.lab.txt").read_text()

pau_ends = []

for l in base_lab.splitlines():
    a, b, c = l.split()
    if c == "pau":
        pau_ends += [float(b)]

pau_ends = pau_ends[:-1]

texts = []
for i, et in enumerate(pau_ends):
    for s in Path(f"output-split-{i}.lab").read_text().splitlines():
        s, e, p = s.split()
        s = str(float(s) + et - 0.1)
        e = str(float(e) + et - 0.1)
        texts += [f"{s}\t{e}\t{p}"]

Path("output-merged.lab").write_text("\n".join(texts))