metadata

license: cc-by-4.0
datasets:
  - bene-ges/spellmapper_en_train_v1
language:
  - en
library_name: nemo

SpellMapper - Spellchecking ASR Customization Model

This model is an alternative to word boosting/shallow fusion approaches:

does not require retraining ASR model;
does not require beam-search/language model (LM);
can be applied on top of any English ASR model output;

Paper: SpellMapper: A non-autoregressive neural spellchecker for ASR customization with candidate retrieval based on n-gram mappings

Documentation page.

How to Use this Model

To use this model you will need to install NVIDIA NeMo.

See Bash-script with example of inference pipeline.

Or play with Tutorial.

Citation

    @inproceedings{inproceedings,
        author = {Antonova, Alexandra and Bakhturina, Evelina and Ginsburg, Boris},
        year = {2023},
        month = {08},
        pages = {1404-1408},
        title = {SpellMapper: A non-autoregressive neural spellchecker for ASR customization with candidate retrieval based on n-gram mappings},
        doi = {10.21437/Interspeech.2023-768}
    }