Hugging Face's logo

MontrealCorpusTools
/
mandarin_mfa

Other
Chinese
montreal-forced-aligner
forced-alignment
Model card Files
xet
 Community

Model Card for mandarin_mfa

This MFA model is for aligning Mandarin speech.

Model Details

Model Description

  • Developed by: Michael McAuliffe
  • Funded by: N/A
  • Model type: Montreal Forced Aligner model
  • Language(s) (NLP): Mandarin
  • License: cc-by-4.0

Uses

Direct Use

This model is intended to be used for forced alignment of Mandarin speech. Please see https://montreal-forced-aligner.readthedocs.io/en/latest/user_guide/troubleshooting.html for details on common fixes. If using Pinyin transcripts, use the G2P models either directly or via the --use_g2p --language unknown flags. Otherwise, MFA will assume Chinese characters and use a tokenizer to split up words and generate pronunciations for unknown words based on their Pinyin conversion.

Out-of-Scope Use

This model cannot provide accurate assessments of goodness of pronunciations or provide transcripts as it is trained to be accepting of variation in pronunciation to provide a reasonable alignment for Mandarin speech.

Bias, Risks, and Limitations

This model will perform best on the variety of speech that it was trained on. The speakers in the training data are all adult speakers, so child speech alignment may not be accurate.

Recommendations

When using this model on a variety that it was not trained on, better results can be attained by adapting the model to the data to be aligned first. See https://montreal-forced-aligner.readthedocs.io/en/latest/user_guide/workflows/adapt_acoustic_model.html and https://github.com/mmcauliffe/mfa-adaptation for example usage and scripts.

How to Get Started with the Model

Use the code below to get started with the model.

To get started, follow the instructions for installing MFA. To align files using this model, use the mfa align command.

Dictionary Details

Details for mandarin_taiwan_mfa dictionary and G2P model

  • Source: wikipron
  • Orthography: Hanzi and Pinyin
  • Phone set: MFA
  • Words: 156,093
  • Phones: 51
  • Graphemes: 18,170
IPA chart
Consonants
Manner Labial Labiodental Alveolar Retroflex Palatal Velar Glottal
Nasal m n n̩ ɲ ŋ ŋ̍
Palatalized mʲ
Stop p t k ʔ
Aspirated pʰ tʰ kʰ
Labialized pʷ tʷ kʷ
Palatalized pʲ tʲ
Affricate ts ʈʂ tɕ
Aspirated tsʰ ʈʂʰ tɕʰ
Labialized tɕʷ
Sibilant s z̩ ʂ ʐ ʐ̩ ɕ
Labialized ɕʷ
Fricative f
Approximant w ɻ j ɥ
Lateral l ʎ
Vowels
Front Near-Front Central Near-Back Back
Close i y u
Close-Mid e ej o ow
ə
Open-Mid
Open a
Diphthongs
  • aw
  • aj
  • ej
  • ow

Details for mandarin_china_mfa dictionary and G2P model

  • Source: wikipron
  • Orthography: Hanzi and Pinyin
  • Phone set: MFA
  • Words: 286,218
  • Phones: 52
  • Graphemes: 18,043
IPA chart
Consonants
Manner Labial Labiodental Alveolar Retroflex Palatal Velar Glottal
Nasal m m̩ n n̩ ɲ ŋ ŋ̍
Palatalized mʲ
Stop p t k ʔ
Aspirated pʰ tʰ kʰ
Labialized pʷ tʷ kʷ
Palatalized pʲ tʲ
Affricate ts ʈʂ tɕ
Aspirated tsʰ ʈʂʰ tɕʰ
Labialized tɕʷ
Sibilant s z̩ ʂ ʐ ʐ̩ ɕ
Labialized ɕʷ
Fricative f
Approximant w ɻ j ɥ
Lateral l ʎ
Vowels
Front Near-Front Central Near-Back Back
Close i y u
Close-Mid e ej o ow
ə
Open-Mid
Open a
Diphthongs
  • aw
  • aj
  • ej
  • ow

Training Details

Training Data

AI-DataTang Corpus

  • Source: https://openslr.org/62/
  • License: CC BY-NC-ND 4.0
  • Dialects: china, erhua
  • Number of hours: 200.38
  • Number of utterances: 237,264
  • Number of speakers: 600
    • Female speakers: 0
    • Male speakers: 0
    • Unknown speakers: 420

Common Voice Chinese (China)

  • Source: https://voice.mozilla.org/en/datasets
  • License: CC-0
  • Dialects: china, erhua
  • Number of hours: 238.57
  • Number of utterances: 169,114
  • Number of speakers: 4,462
    • Female speakers: 152
    • Male speakers: 748
    • Unknown speakers: 3,562

Common Voice Chinese (Taiwan)

  • Source: https://voice.mozilla.org/en/datasets
  • License: CC-0
  • Dialects: taiwan
  • Number of hours: 72.75
  • Number of utterances: 82,137
  • Number of speakers: 1,706
    • Female speakers: 191
    • Male speakers: 458
    • Unknown speakers: 1,054

AISHELL-3

  • Source: https://openslr.org/93/
  • License: Apache 2.0
  • Dialects: china, erhua
  • Number of hours: 160.88
  • Number of utterances: 127,274
  • Number of speakers: 360
    • Female speakers: 0
    • Male speakers: 0
    • Unknown speakers: 360

THCHS-30

  • Source: https://openslr.org/18/
  • License: Apache 2.0
  • Dialects: china, erhua
  • Number of hours: 34.16
  • Number of utterances: 13,387
  • Number of speakers: 60
    • Female speakers: 0
    • Male speakers: 0
    • Unknown speakers: 60

GlobalPhone Chinese-Mandarin

Training Procedure

Preprocessing

Preprocessing include fixes and orthographic standardization to various corpora.

Training Hyperparameters

Evaluation

Testing Data, Factors & Metrics

Testing Data

N/A

Factors

N/A

Metrics

N/A

Results

N/A

Summary

Technical Specifications

Model Architecture and Objective

HMM-GMM model

Software

This model was trained via the Montreal Forced Aligner.

Citation

BibTeX:

@techreport{mfa_mandarin_mfa_acoustic_2026,
    author={McAuliffe, Michael and Sonderegger, Morgan},
    title={Mandarin MFA acoustic model v3.3.0},
    address={\url{https://huggingface.co/MontrealCorpusTools/mandarin_mfa}},
    year={2026},
    month={Jun},
}

APA:

McAuliffe, M. & Sonderegger, M. (2026). Mandarin MFA acoustic model v3.3.0. Available at https://huggingface.co/MontrealCorpusTools/mandarin_mfa.

Contact

For questions and issues, please file an issue either for this model at https://huggingface.co/MontrealCorpusTools/mandarin_mfa/discussions or for larger MFA issues at https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner/issues.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
Other
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support