Abstract
A persistent challenge in noisy-channel decoding is the recognition of unusual or context-specific inputs such as named entities, e.g., names of personal contacts, songs, apps; specialized phrases, e.g., voice commands, technical/medical/legal jargon, etc. Noisy-channel decoding is a problem that occurs in automatic speech recognizers (ASR). This disclosure describes techniques to improve lattice diversity in noisy-channel decoding, resulting in richer decoding lattices. Decoding errors are reduced by using the decoder to model uncertain/unknown entities, e.g., by embedding uncertainty into the training data. Primary errors such as misrecognition of named entities become easier to fix because secondary errors are reduced or avoided. The resulting decoder is robust to words and phrases that are missing or uncommon in training data.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Velikovich, Leonid and Aleksic, Petar, "Decoders that Model Unknown Entities", Technical Disclosure Commons, (March 09, 2022)
https://www.tdcommons.org/dpubs_series/4954