Montreal Forced Aligner for linux-aarch64

A self-contained Docker build of Montreal Forced Aligner 3.3.9 for linux-aarch64 such as the NVIDIA DGX Spark, Apple Silicon, Ampere and AWS Graviton. conda-forge publishes MFA only where all of its native dependencies have aarch64 builds, and four of them do not, so mamba install montreal-forced-aligner fails on ARM Linux out of the box. This image supplies those four itself - three built from source into a local channel, sox from apt - and resolves the rest from conda-forge normally.

Unofficial build. A community packaging for aarch64, not affiliated with or endorsed by the Montreal Forced Aligner project or its authors.

Usage

Pull the prebuilt image from Docker Hub. It is an arm64 image; on a non-arm host add --platform linux/arm64 to run it under QEMU.

docker pull grammatek/mfa_arm64

Or build it yourself:

# 1. Fetch the OpenGrm source tarballs. They are SHA-256 verified.
./fetch-sources.sh

# 2a. Build natively on arm64:
docker build -t grammatek/mfa_arm64 .
# 2b. or cross-build from an amd64 host under QEMU, which is slower; the kalpy compile alone takes about an hour.
docker buildx build --platform linux/arm64 -t grammatek/mfa_arm64 --load .

Align a corpus against a pronunciation dictionary and acoustic model. This assumes the current host directory holds:

• corpus/ - speaker subfolders of audio, each file paired with a .lab transcript
• corpus.dict - the pronunciation dictionary
• acoustic.zip - the acoustic model

-v "$PWD:/data" mounts that directory at /data, so the /data/... arguments resolve to these local files and the alignment is written back to ./out. --user keeps the output owned by you, and MFA_ROOT_DIR and HOME point at the writable mount.

docker run --rm -v "$PWD:/data" \
  -e MFA_ROOT_DIR=/data/mfa_root -e HOME=/data \
  --user "$(id -u):$(id -g)" \
  grammatek/mfa_arm64 \
  align /data/corpus /data/corpus.dict /data/acoustic.zip /data/out \
  -j "$(nproc)" --clean --output_format json

The entrypoint is mfa, so any mfa subcommand works.

The four gaps

conda-forge ships no aarch64 build for four of MFA's dependencies:

• baumwelch and ngram - built from their feedstocks. fetch-sources.sh pulls the OpenGrm tarballs because the recipe's own download URL is dead.
• kalpy - pinned to 0.9.0. MFA 3.3.9 calls its TrainingGraphCompiler(use_g2p=...), which later releases dropped. 0.9.0 needs a one-line patch to compile on the aarch64 gcc, kalpy_090/fix-lattice-return.patch.
• sox - installed from Ubuntu apt and removed from MFA's conda dependencies.

Layout

Dockerfile                                          the whole build
fetch-sources.sh                                    downloads and SHA-verifies the OpenGrm tarballs
sources/                                            populated by fetch-sources.sh
  baumwelch-0.3.11.tar.gz
  ngram-1.3.17.tar.gz
patches/                                            feedstock patches
  baumwelch-feedstock__local-source.patch           point the source url at the fetched tarball
  ngram-feedstock__local-source.patch               same for ngram
kalpy_090/
  fix-lattice-return.patch                          C++ backport so kalpy 0.9.0 compiles

Pinned versions

• montreal-forced-aligner 3.3.9
• kalpy 0.9.0
• kaldi 5.5.1172 cpu
• openfst 1.8.4
• baumwelch 0.3.11
• ngram 1.3.17

These match the set shipped by the official amd64 image.

Updating

conda-forge is a moving target: the feedstock default branches advance and the channel resolves the rest of the dependency tree live. Every input here is therefore pinned - the base image by digest, the four feedstocks by commit, and conda-build, conda-forge-pinning, the package versions and source hashes by value. To move to a newer MFA:

• Re-pin MFA_FEEDSTOCK to a specific montreal-forced-aligner-feedstock commit and set MFA_VERSION to match it. Do not track the branch head. When this was written the feedstock had already advanced from build number 2 to 3 and relaxed its kalpy pin, the kind of silent drift the pin prevents.
• Confirm the recipe edits still fit that commit: one number: line to bump, a sox run dependency to drop, and a kalpy pin the built version satisfies.
• If the new MFA needs a different kalpy, set KALPY_VERSION and KALPY_SHA256 from the kalpy-kaldi release on PyPI, and check whether kalpy_090/fix-lattice-return.patch still applies or is even needed. Later kalpy dropped the use_g2p API MFA calls, so the version is not a free choice.
• Refresh the base image digest and the other feedstock commits the same way.

After any bump, run a --no-cache build and the align check. Pinned inputs keep rebuilds compatible, but the transitive conda-forge solve is not bit-identical, so only a real build confirms it still works.

References

• Montreal Forced Aligner - documentation and source
• Official Docker images, amd64 only - Docker Hub and installation docs
• Training an acoustic model - MFA user guide
• MFA 3.3.9 - release notes