English (US) ARPA acoustic model v2.0.0#
@techreport{mfa_english_us_arpa_acoustic_2022,
author={McAuliffe, Michael and Sonderegger, Morgan},
title={English (US) ARPA acoustic model v2.0.0},
address={\url{https://mfa-models.readthedocs.io/acoustic/English/English (US) ARPA acoustic model v2_0_0.html}},
year={2022},
month={Mar},
}
Training corpora
Pronunciation dictionaries |
 |
Installation#
Install from the MFA command line:
mfa model download acoustic english_us_arpa
Or download from the release page.
Intended use#
This model is intended for forced alignment of English transcripts.
This model uses the ARPA phone set for English, and was trained with the pronunciation dictionaries above. Pronunciations can be added on top of the dictionary, as long as no additional phones are introduced.
Performance Factors#
As forced alignment is a relatively well-constrained problem (given accurate transcripts), this model should be applicable to a range of recording conditions and speakers. However, please note that it was trained on read speech in low-noise environments, so as your data diverges from that, you may run into alignment issues or need to increase the beam size of MFA or see other recommendations in the troubleshooting section below.
Please note as well that MFA does not use state-of-the-art ASR models for forced alignment. You may get better performance (especially on speech-to-text tasks) using other frameworks like Coqui.
Metrics#
Acoustic models are typically generated as one component of a larger ASR system where the metric is word error rate (WER). For forced alignment, there is typically not the same sort of gold standard measure for most languages.
As a rough approximation of the acoustic model quality, we evaluated it against the corpus it was trained on alongside a language model trained from the same data. Key caveat here is that this is not a typical WER measure on held out data, so it should not be taken as a hard measure of how well an acoustic model will generalize to your data, but rather is more of a sanity check that the training data quality was sufficiently high.
Using the pronunciation dictionaries and language models above:
WER:
0%CER:
0%
Ethical considerations#
Deploying any Speech-to-Text model 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 STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this 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.
Troubleshooting issues#
Machine learning models (like this acoustic model) perform best on data that is similar to the data on which they were trained.
The primary sources of variability in forced alignment will be the applicability of the pronunciation dictionary and how similar the speech, demographics, and recording conditions are. If you encounter issues in alignment, there are couple of avenues to improve performance:
-
MFA defaults to a narrow beam to ensure quick alignment and also as a way to detect potential issues in your dataset, but depending on your data, you might benefit from boosting the beam to 100 or higher.
Add pronunciations to the pronunciation dictionary
This model was trained a particular dialect/style, and so adding pronunciations more representative of the variety spoken in your dataset will help alignment.
Check the quality of your data
MFA includes a validator utility, which aims to detect issues in the dataset.
Use MFA’s anchor utility to visually inspect your data as MFA sees it and correct issues in transcription or OOV items.
Adapt the model to your data
MFA has an adaptation command to adapt some of the model to your data based on an initial alignment, and then run another alignment with the adapted model.