Google recently announced the general language model research that was put into use in November last year. The results show that the Google language model can now recognize more than 100 languages. In some general language recognition, the semantic understanding performance of the Google language model is even more impressive. Compared with Whisper, a large-scale language model launched by OpenAI, the semantic recognition error rate is lower.
According to Google’s official introduction, the Google language model can carry out continuous self-learning, and can continuously modify the overall structure of the language model with the blessing of the BEST-RQ algorithm, so as to complete operations such as continuous analysis and learning of language structure. In addition, when the Google language model performs semantic understanding, it also mobilizes text injection and supervised loss functions, so that the semantic understanding of the language model is more accurate.
At present, when the Google language model is processing more than 70 language translations, the error rate can be kept within 30%.
It is worth mentioning that the Google language model is even more accurate than Whisper in dealing with the spoken English CORAAL used by African-Americans, SpeechStew with mixed accents, and FLEURS in other languages.