Deliberation of transducer-based streaming for cascade encoder
The transducer-based cascaded encoder model addresses the latency and accuracy challenges of ASR by using a two-pass transducer decoder system, enabling low-latency, streaming speech recognition with improved accuracy for real-time applications.
JP2024175059AActive Publication Date: 2024-12-17GOOGLE LLC
Patent Information
- Application Number
- JP2024160405
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-10-04
- Filing Date
- 2024-09-17
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2042-09-19
Smart Images
Figure 2024175059000001_ABST
Abstract
To provide a transducer-based deliberation model of a cascade encoder and a mounting method of the model.SOLUTION: Automatic voice recognition 200 includes: a step for receiving a sequence of acoustic frames; a step for generating a first high order feature representation 212 on a corresponding acoustic frame in the sequence by a first encoder; a step for generating a first pass speech recognition hypothesis 120a on the corresponding first high order feature representation by a first pass transducer decoder 201; a step for generating text encoding 242 on the corresponding first pass speech recognition hypothesis by a text encoder; a step for generating a second high order feature representation 222 on the corresponding first high order feature representation by a second encoder; and a step for generating a second pass speech hypothesis 120b by using the second high order feature representation by a second pass transducer decoder 202 and corresponding text encoding.SELECTED DRAWING: Figure 2
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Deliberation Model-Based Two-Pass End-To-End Speech Recognition
US20210225369A1