Method and system for performing domain adaptation of end-to-end automatic speech recognition model
US11315548B1 · kind B1 · utility
Assignee
Inventors
Key dates
| Filing date | Aug 30, 2021 |
| Grant date | Apr 26, 2022 |
| Priority date | — |
| Expiry date | Aug 30, 2041 |
Classification
- Technology area (CPC G)Physics
- CPC primaryG10L15/22
- WIPO fieldComputer technology
- WIPO sectorElectrical engineering
Abstract
A method for performing domain adaptation of an end-to-end (E2E) automatic speech recognition (ASR) model. The method includes: obtaining an un-adapted version of the E2E ASR model trained using a first set of transcriptions, the un-adapted version of E2E ASR model including an encoder network, a first prediction network and a joint network; using the first set of transcriptions, while keeping parameters of first prediction network fixed, to train a language model output component to match the first prediction network; using a second set of transcriptions, while keeping parameters of language model output component fixed, to fine-tune the first prediction network for obtaining a second prediction network; and generating an adapted version of the E2E ASR model, wherein the adapted version of the E2E ASR model comprises the encoder network, the second prediction network, the language model output component, and the joint network.
Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.