On Semi-Supervised LF-MMI Training of Acoustic Models with Limited Data

Imran Sheikh, Emmanuel Vincent, Irina Illina


This work investigates semi-supervised training of acoustic models (AM) with the lattice-free maximum mutual information (LF-MMI) objective in practically relevant scenarios with a limited amount of labeled in-domain data. An error detection driven semi-supervised AM training approach is proposed, in which an error detector controls the hypothesized transcriptions or lattices used as LF-MMI training targets on additional unlabeled data. Under this approach, our first method uses a single error-tagged hypothesis whereas our second method uses a modified supervision lattice. These methods are evaluated and compared with existing semi-supervised AM training methods in three different matched or mismatched, limited data setups. Word error recovery rates of 28 to 89% are reported.


 DOI: 10.21437/Interspeech.2020-2242

Cite as: Sheikh, I., Vincent, E., Illina, I. (2020) On Semi-Supervised LF-MMI Training of Acoustic Models with Limited Data. Proc. Interspeech 2020, 986-990, DOI: 10.21437/Interspeech.2020-2242.


@inproceedings{Sheikh2020,
  author={Imran Sheikh and Emmanuel Vincent and Irina Illina},
  title={{On Semi-Supervised LF-MMI Training of Acoustic Models with Limited Data}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={986--990},
  doi={10.21437/Interspeech.2020-2242},
  url={http://dx.doi.org/10.21437/Interspeech.2020-2242}
}