Pretrained models

Pretrained acoustic models

As part of using the Montreal Forced Aligner in our own research, we have trained acoustic models for a number of languages. If you would like to use them, please download them below. Please note the dictionary that they were trained with to see more information about the phone set. When using these with a pronunciation dictionary, the phone sets must be compatible. If the orthography of the language is transparent, it is likely that we have a G2P model that can be used to generate the necessary pronunciation dictionary.

Language	Link	Corpus	Phone set
Arabic	Not available yet	GlobalPhone	GlobalPhone
Bulgarian	Bulgarian acoustic model	GlobalPhone	GlobalPhone
Croatian	Croatian acoustic model	GlobalPhone	GlobalPhone
Czech	Czech acoustic model	GlobalPhone	GlobalPhone
English	English acoustic model	LibriSpeech	Arpabet
French (FR)	French (FR) acoustic model	GlobalPhone	GlobalPhone
French (FR)	French (Prosodylab) acoustic model	GlobalPhone	Prosodylab [1]
French (QC)	French (QC) acoustic model	Lab speech	Prosodylab [1]
German	German acoustic model	GlobalPhone	GlobalPhone
German	German (Prosodylab) acoustic model	GlobalPhone	Prosodylab [3]
Hausa	Hausa acoustic model	GlobalPhone	GlobalPhone
Japanese	Not available yet	GlobalPhone	GlobalPhone
Korean	Korean acoustic model	GlobalPhone	GlobalPhone
Mandarin	Mandarin acoustic model	GlobalPhone	GlobalPhone
Polish	Polish acoustic model	GlobalPhone	GlobalPhone
Portuguese	Portuguese acoustic model	GlobalPhone	GlobalPhone
Russian	Russian acoustic model	GlobalPhone	GlobalPhone
Swahili	Swahili acoustic model	GlobalPhone	GlobalPhone
Swedish	Swedish acoustic model	GlobalPhone	GlobalPhone
Tamil	Not available yet	GlobalPhone	GlobalPhone
Thai	Thai acoustic model	GlobalPhone	GlobalPhone
Turkish	Turkish acoustic model	GlobalPhone	GlobalPhone
Ukrainian	Ukrainian acoustic model	GlobalPhone	GlobalPhone
Vietnamese	Vietnamese acoustic model	GlobalPhone	GlobalPhone
Wu	Not available yet	GlobalPhone	GlobalPhone

Pretrained G2P models

Included with MFA is a separate tool to generate a dictionary from a preexisting model. This should be used if you’re aligning a dataset for which you have no pronunciation dictionary or the orthography is very transparent. We have pretrained models for several languages, which can be downloaded below. These models were generated using Phonetisaurus Phonetisaurus and the GlobalPhone dataset. This means that they will only work for transcriptions which use the same alphabet. Current language options are: Arabic, Bulgarian, Mandarin, Czech, Polish, Russian, Swahili, Ukrainian, and Vietnamese, with the following accuracies when trained on 90% of the data and tested on 10%:

Language	Link	Accuracy	Orthography system	Phone set
Arabic	Arabic G2P model	95.4	Romanized [1]	GlobalPhone
Bulgarian	Bulgarian G2P model	97.3	Cyrillic alphabet	GlobalPhone
Croatian	Croatian G2P model	92.7	Latin alphabet	GlobalPhone
Czech	Czech G2P model	96.8	Latin alphabet	GlobalPhone
French	French G2P model	93.2	Latin alphabet	GlobalPhone
French	French (Prosodylab) G2P model [1]	95.2	Latin alphabet	Prosodylab
German	German G2P model	67.0	Latin alphabet	GlobalPhone
German	German (Prosodylab) G2P model [3]	94.1	Latin alphabet	Prosodylab
Hausa	Hausa G2P model	70.1	Latin alphabet	GlobalPhone
Japanese	Japanese G2P model	82.1	Romanized	GlobalPhone
Korean	Korean G2P model	89.5	Hangul	GlobalPhone
Mandarin	Mandarin Pinyin G2P model	99.9	Pinyin	Pinyin phones
Mandarin	Mandarin Character G2P model [4]	83.2	Hanzi	Pinyin phones
Polish	Polish G2P model	98.8	Latin alphabet	GlobalPhone
Portuguese	Portuguese G2P model	86.5	Latin alphabet	GlobalPhone
Russian	Russian G2P model	96.4	Cyrillic alphabet	GlobalPhone
Spanish	Spanish G2P model	94.0	Latin alphabet	GlobalPhone
Swahili	Swahili G2P model	99.9	Latin alphabet	GlobalPhone
Swedish	Swedish G2P model	83.3	Latin alphabet	GlobalPhone
Thai	Thai G2P model	71.7	Thai script	GlobalPhone
Turkish	Turkish G2P model	83.3	Latin alphabet	GlobalPhone
Ukrainian	Ukrainian G2P model	98.0	Cyrillic alphabet	GlobalPhone
Vietnamese	Vietnamese G2P model	98.2	Vietnamese alphabet	GlobalPhone
Wu	Wu G2P model [5]	77.5	Hanzi	GlobalPhone

[1] (1,2,3,4)

The ProsodyLab French dictionary is based on Lexique with substitutions for numbers and special characters. Note that Lexique is known to currently not work with the aligner, see the Github issue for more information and status.

[2]

Please see the GlobalPhone documentation for how the romanization was done for Arabic.

[3] (1,2)

The German dictionary used in training is available in the ProsodyLab dictionary repository. See http://www.let.uu.nl/~Hugo.Quene/personal/phonchar.html for more information on the CELEX phone set for German and how it maps to other phonesets.

[4]

The Mandarin character dictionary that served as the training data for this model was built by mapping between characters in .trl files and pinyin syllables in .rmn files in the GlobalPhone corpus.

[5]

The Wu G2P model was trained a fairly small lexicon, so it likely does not have the coverage to be a robust model for most purposes. Please check carefully any resulting dictionaries, as they are likely to have missing syllables from from unknown symbols.