What’s new in 2.1

Version 2.1 of the Montreal Forced Aligner changes the command line interface to use click, transition fully to use postgresql as the backend, expand functionality for segmentation and speaker diarization, and support the latest alpha release of Anchor, along with other improvements like fine-tuning alignments and generating phone-level confidences. See 2.1 Changelog for a more specific changes.

Click

MFA 2.1 uses click instead of Python’s default argparse. The primary benefit for this is in better validation of arguments and the ability to prompt the user if they forget an argument like dictionary_path.

Dependency on postgresql

In 2.0, the default database backend was SQLite, which allowed for rapid development, but it lacks more advanced functionality in other database backends. PostgreSQL was supported as well, but it requires a persistent running server rather a single database file for SQLite. As more advanced functionality has been shifted to SQL over Python to speed up querying and processes, it made sense to drop SQLite support in favor of pure PostgreSQL.

Segmentation with SpeechBrain

In addition to the simple energy-based VAD used in Kaldi, MFA is capable of using SpeechBrain’s VAD model to generate better segmentation of long audio files. See Segment transcribed files (mfa segment) for more information.

Note

SpeechBrain is not installed by default with conda install montreal-forced-aligner, so please refer to Installation for more details.

Speaker diarization

Speaker diarization has been overhauled to make use of pretrained ivector models or SpeechBrain’s https://speechbrain.readthedocs.io/en/latest/API/speechbrain.pretrained.interfaces.html#speechbrain.pretrained.interfaces.SpeakerRecognition. Additionally, with the Dependency on postgresql in MFA 2.1, we have integrated with pgvector, which allows for storage and querying of utterance and speaker ivectors.

Note

SpeechBrain is not installed by default with conda install montreal-forced-aligner, so please refer to Installation for more details.

Anchor annotator GUI

See also Anchor annotator (mfa anchor) for more information on using the annotation GUI.