Montreal Forced Aligner Dockerfile

Dockerfile

# FROM kaldiasr/kaldi:latest
FROM python:3.6-buster

WORKDIR "/workspace"

RUN apt-get clean \
        && apt-get update \
        && apt-get install -y wget \
          automake autoconf sox libtool subversion libatlas3-base gfortran \
        && apt-get -y autoremove

# ADD requirements.txt requirements.txt
# RUN pip install -r requirements.txt

# kaldi (specific version!)
ENV KALDI_VERSION=094d22746b604fd20c2b8730966c9d0bc9f2170b
RUN cd /opt && \
  git clone https://github.com/kaldi-asr/kaldi.git && \
  cd kaldi && \
  git checkout $KALDI_VERSION && \
  cd tools && \
  make -j $(NPROC) && \
  make openblas && \
  cd ../src && \
  ./configure --shared --openblas-root=/opt/kaldi/tools/OpenBLAS/install && \
  make depend -j $(NPROC) && \
  make -j 4


# OpenGrm-Ngram
ENV OPENGRM_NGRAM=1.3.4
RUN cd /opt && \
  wget "http://www.opengrm.org/twiki/pub/GRM/NGramDownload/opengrm-ngram-$OPENGRM_NGRAM.tar.gz" && \
  tar -xf "opengrm-ngram-$OPENGRM_NGRAM.tar.gz" && \
  rm "opengrm-ngram-$OPENGRM_NGRAM.tar.gz" && \
  cd "opengrm-ngram-$OPENGRM_NGRAM" && \
  export LD_LIBRARY_PATH=/opt/kaldi/tools/openfst/lib && \
  export CPLUS_INCLUDE_PATH=/opt/kaldi/tools/openfst/src/include && \
  ./configure --prefix=`pwd`/install && \
  make -j 4 && \
  make install

# Phonetisaurus
ENV PHONETISAURUS_VERSION=64719ca40c17cb70d810fffadac52c97984ca539
RUN cd /opt && \
  git clone https://github.com/AdolfVonKleist/Phonetisaurus.git && \
  cd Phonetisaurus && \
  git checkout $PHONETISAURUS_VERSION && \
  ./configure \
    --enable-static=no \
    --with-openfst-includes=/opt/kaldi/tools/openfst/include \
    --with-openfst-libs=/opt/kaldi/tools/openfst/lib && \
    make -j 4

# aligner
ENV MONTREAL_ALIGNER_VERSION=1.0.1
RUN cd /opt && \
  git clone https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner.git && \
  cd Montreal-Forced-Aligner && \
  python thirdparty/kaldi_binaries.py /opt/kaldi && \
  python thirdparty/opengrm_ngram_binaries.py /opt/opengrm-ngram-$OPENGRM_NGRAM && \
  python thirdparty/phonetisaurus_binaries.py /opt/Phonetisaurus && \
  pip install -r requirements.txt && \
  mkdir pretrained_models && \
  cd pretrained_models && \
  wget https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner/raw/dc09bb3d7302bc66eb8fdef543b44aa0fab61b07/pretrained_models/english.zip && \
  cd ../ && \
  python freezing/freeze.py

WORKDIR "/opt/Montreal-Forced-Aligner/dist/montreal-forced-aligner"

# only requirement is TextGridTools
RUN pip install tgt

CMD ["/bin/bash"]

textgrid_to_alignments.py

# use a different docker image!
# make build_align && make run_align
# bin/mfa_align \
#   /datasets/CommonVoice/en/speakers \
#   /datasets/slr60/english.dict \
#   /opt/Montreal-Forced-Aligner/dist/montreal-forced-aligner/pretrained_models/english.zip \
#   /output/montreal-aligned/cv-en/
# bin/mfa_validate_dataset \
#   /datasets/slr60/test-clean \
#   /datasets/slr60/english.dict\
#   english

import sys
import tgt
from pathlib import Path
from tqdm import tqdm

DATASET = 'dev-clean'
# DATASET = 'train-clean-100'
# DATASET = 'train-clean-360'
dataset_path = Path('/datasets/slr60/{}'.format(DATASET))
base_path = Path('/output/montreal-aligned/{}'.format(DATASET))
speaker_dirs = [f for f in base_path.glob("*") if f.is_dir()]

for speaker_dir in tqdm(speaker_dirs):
    book_dirs = [f for f in speaker_dir.glob("*") if f.is_dir()]

    for book_dir in book_dirs:
        alignment_file = dataset_path.joinpath(
            speaker_dir.stem,
            book_dir.stem,
            "{0}_{1}.alignment.txt".format(speaker_dir.stem, book_dir.stem)
        )

        with open(alignment_file, 'w', encoding='utf-8') as out_file:
            # find our textgrid files
            textgrid_files = sorted([f for f in book_dir.glob("*.TextGrid") if f.is_file()])

            # process each grid file and add to our output
            for textgrid_file in textgrid_files:
                # read the raw transcript as well
                transcript_file = dataset_path.joinpath(
                    speaker_dir.stem,
                    book_dir.stem,
                    "{0}.txt".format(textgrid_file.stem)
                )
                with open(transcript_file, 'r', encoding='utf-8') as in_file:
                    transcript = in_file.read()

                # read the grid
                input = tgt.io.read_textgrid(textgrid_file)
                print("input: {}".format(input))
                sys.exit(1)

                # get all the word tiers
                word_tier = input.get_tier_by_name('words')

                out_file.write("{0} \"{1}\" \"{2}\" {3}\n".format(
                    textgrid_file.stem,
                    ",".join(list(map(lambda interval: interval.text, word_tier.intervals))),
                    ",".join(list(map(lambda interval: str(interval.end_time), word_tier.intervals))),
                    transcript
                ))