Voice conversion method and device, electronic equipment and storage medium
By generating acoustic features through posterior probability features of speech and historical memory information, the limitations of multi-source speaker conversion in existing technologies are overcome, enabling low-cost, real-time speech conversion that is suitable for scenarios such as live streaming and instant messaging.
CN115547349BActive Publication Date: 2026-07-21出门问问(苏州)信息科技有限公司
View PDF 2 Cites 0 Cited by
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 出门问问(苏州)信息科技有限公司
- Filing Date
- 2022-09-19
- Publication Date
- 2026-07-21
AI Technical Summary
Technical Problem
Existing speech conversion technologies require parallel corpus training, which limits the application of speech conversion from multiple source speakers to the same target speaker, and is difficult to apply in real-time scenarios, resulting in high costs.
Method used
Acoustic features are generated using posterior probability features and historical memory information. Speech blocks are received via WebSocket and converted in real time using a pre-trained acoustic model, achieving many-to-one speech conversion.
Benefits of technology
It achieves low-cost, real-time speech conversion, supports conversion from multiple source speakers to the same target speaker, and is suitable for scenarios such as live streaming and instant messaging.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN115547349B_ABST
Abstract
The present disclosure provides a voice conversion method and device, electronic equipment and storage medium. The voice conversion method of the present disclosure comprises: obtaining a current source speech block; extracting a speech posterior probability feature of the current source speech block; generating an acoustic feature corresponding to the current source speech block by using the speech posterior probability feature of the current source speech block and historical memory information of a previously obtained previous source speech block; and synthesizing the acoustic feature corresponding to the current source speech block into a target speech block; wherein the target speech block has a timbre feature of a target speaker, and a speaker of the source speech block is a source speaker. The present disclosure can realize real-time voice conversion, has low cost, fast execution speed, and supports one-to-many voice conversion.
Need to check novelty before this filing date? Find Prior Art