What is forced alignment?¶
Forced alignment is a technique to take an orthographic transcription of an audio file and generate a time-aligned version using a pronunciation dictionary to look up phones for words.
Many languages have pretrained acoustic models available for download and use (Pretrained acoustic models)
Montreal Forced Aligner¶
Pipeline of training¶
The Montreal Forced Aligner goes through three stages of training. The first pass of alignment uses monophone models, where each phone is modelled the same regardless of phonological context. The second pass uses triphone models, where context on either side of a phone is taken into account for acoustic models. The final pass enhances the triphone model by taking into account speaker differences, and calculates a transformation of the mel frequency cepstrum coefficients (MFCC) features for each speaker.
If you run into any issues, please check the mailing list for fixes/workarounds or to post a new issue.
Use of speaker information¶
A key feature of the Montreal Forced Aligner is the use of speaker adaptatation in alignment. The command line interface provides multiple ways of grouping audio files by speaker, depending on the input file format (either Prosodylab-aligner format or TextGrid format). In addition to speaker-adaptation in the final pass of alignment, speaker information is used for grouping audio files together for multiprocessing and ceptstral mean and variance normalization (CMVN). If speakers are not properly specified, then feature calculation might not succeed due to limits on the numbers of files open.
Other forced alignment tools¶
Most tools for forced alignment used by linguists rely on the HMM Toolkit (HTK; HTK homepage), including:
- Prosodylab-aligner (Prosodylab-aligner homepage)
- Penn Phonetics Forced Aligner (P2FA, P2FA homepage)
- FAVE-align (FAVE-align homepage)
- (Web) MAUS(MAUS homepage)
Montreal Forced Aligner is most similar to the Prosodylab-aligner, and was developed at the same lab. Because the Montreal Forced Aligner uses a different toolkit to do alignment, trained models cannot be used with the Prosodylab-aligner, and vice versa.
Another Kaldi-based forced aligner is Gentle (Gentle homepage) which uses Kaldi’s neural networks to align English data. The Montreal Forced Aligner allows for training on any data that you might have, and can be used with languages other than English.
McAuliffe, Michael, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger (2017). Montreal Forced Aligner [Computer program]. Version 0.9.0, retrieved 17 January 2017 from http://montrealcorpustools.github.io/Montreal-Forced-Aligner/.
McAuliffe, Michael, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger (2017).
Montreal Forced Aligner: an accurate and trainable aligner using Kaldi. Presented at the 91st Annual Meeting of the
Linguistic Society of America, Austin, TX.