Japanese tokenizer v2.2.0#

Maintainer: Montreal Forced Aligner
Language: Japanese
Model type: Tokenizer
Architecture: pynini
Model version: v2.2.0
Trained date: 2023-02-15
Compatible MFA version: v2.2.0
License: CC BY 4.0
Citation:

@techreport{mfa_japanese_mfa_tokenizer_2023,
	author={McAuliffe, Michael and Sonderegger, Morgan},
	title={Japanese tokenizer v2.2.0},
	address={\url{https://mfa-models.readthedocs.io/tokenizer/Japanese/Japanese tokenizer v2_2_0.html}},
	year={2023},
	month={Feb},
}

If you have comments or questions about this model, you can check previous MFA model discussion posts or create a new one.

Installation#

Install from the MFA command line:

mfa model download tokenizer japanese_mfa

Or download from the release page.

Intended use#

This model is intended for very basic tokenization of Japanese transcripts.

These tokenization models just use an ngram model for predicting word boundaries based on input text, and are provided only for completeness and in the off chance that they’re useful in bootstrapping corpus development.

Performance Factors#

Third party dedicated tokenizers are likely to be more robust, but this tokenizer is more likely to work well with MFA dictionaries and acoustic models.

Metrics#

The model was trained on 90% of the dictionary and evaluated on a held-out 10% and evaluated with utterance error rate and character error rate.

Training data#

This model was trained on the following data set:

Utterances: 1,482,887
Graphemes: 3,412

Evaluation#

This model was evaluated on the following data set:

Utterances: 164,765
UER: 99.98%
CER: 57.23%

Ethical considerations#

Deploying any model involving language into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias#

You should assume every machine learning model has demographic bias unless proven otherwise. For this tokenizer, this model was trained on a very specific subset of Japanese at the time it was collected that will typically not represent spontaneous speech in the current time. Do not use this model in production, but if you do so, you should acknowledge bias as a potential issue.

Surveillance#

Speech-to-Text technologies may be misused to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in many countries. You should not assume consent to record and analyze private speech.