A speech translation and speech recognition method based on sequence dynamic compression
By using a sequence-based dynamic compression method to adaptively compress speech data, the problems of excessive computational resources and information loss in existing technologies are solved, thereby improving the efficiency and accuracy of speech translation and speech recognition.
CN116206616BActive Publication Date: 2026-07-03XIAONIU FANYI
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAONIU FANYI
- Filing Date
- 2022-12-30
- Publication Date
- 2026-07-03
AI Technical Summary
Technical Problem
Existing speech translation and speech recognition methods suffer from excessive computational resources or information loss when compressing speech data.
Method used
A sequence-based dynamic compression method is adopted, which uses an acoustic encoder and a text encoder to perform length prediction, dynamic compression, feature fusion and encoding of frame-level feature sequences. Adaptive compression is performed using a multi-layer Transformer structure and convolutional layers to dynamically adjust the compression ratio of each sample.
Benefits of technology
This effectively reduces the demand for computing resources, avoids information loss, and improves the coding efficiency of acoustic models and the performance of the final model.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN116206616B_ABST
Abstract
This invention relates to a speech translation and speech recognition method based on dynamic sequence compression, belonging to the field of natural language processing technology. It solves the problems of existing speech translation or speech recognition methods being unable to effectively compress speech data, leading to excessive computational resources; or being unable to perform progressive dynamic compression of speech data, resulting in over-compression and information loss. The speech translation method of this invention includes: acquiring source language speech data to be translated; performing length prediction, dynamic compression, feature fusion, and encoding on the feature sequence of the speech data using an acoustic encoder to obtain an acoustic encoder latent vector; using a text encoder to perform text modality conversion on the acoustic encoder latent vector and extract and encode features to obtain a text encoder latent vector; and inputting the acoustic encoder latent vector and the text encoder latent vector into a decoder for decoding to obtain the target language translation text corresponding to the source language speech data.
Need to check novelty before this filing date? Find Prior Art