Speech Library Expansion for Coherent Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech synthesis technologies, particularly in star and personalized scenarios, face challenges with coherence and naturalness due to limited speech data, making it impractical to record large-scale corpora, which results in incoherent synthesis effects.
Innovation Solution
Expanding a speech library using a pre-trained speech synthesis model and synthesized text, incorporating manually-collected original materials and updating the library with synthesized speech, allowing for more extensive language materials and improved synthesis quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large-scale speech corpus is recorded to improve synthesis quality, then the coherence and naturalness of speech synthesis is improved, but the cost and time required for data collection increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-training a speech synthesis model using available speech data before performing the actual speech splicing and synthesis. This pre-trained model is then used to generate synthesized speech that expands the limited speech library, enabling high-quality synthesis without requiring extensive manual data collection at the time of use.
Solution Approach 2:
The patent uses copying by generating synthesized speech copies from the pre-trained model and adding them to the speech library as synthesized language materials. These synthesized speech segments serve as virtual copies that expand the effective size of the speech library without requiring additional physical recording sessions.
2Reliability
If a large-scale speech corpus is recorded to improve synthesis quality, then the coherence and naturalness of speech synthesis is improved, but the cost of recording increases
Solution Approach 1:
The patent uses copying by generating synthesized speech copies from the pre-trained model and adding them to the speech library as synthesized language materials. These synthesized speech segments serve as virtual copies that expand the effective size of the speech library without requiring additional physical recording sessions.
Solution Approach 2:
The system applies self-service by using the pre-trained speech synthesis model to automatically generate the necessary speech data. The model serves itself by producing synthesized speech that expands the library, eliminating the need for external human recorders and manual data collection processes.
3Productivity
If only a small amount of speech data is collected to reduce cost and time, then the cost and time are reduced, but the speech synthesis effect becomes incoherent and unnatural
Solution Approach 1:
The patent applies preliminary action by pre-training a speech synthesis model using available speech data before performing the actual speech splicing and synthesis. This pre-trained model is then used to generate synthesized speech that expands the limited speech library, enabling high-quality synthesis without requiring extensive manual data collection at the time of use.
Solution Approach 2:
The patent applies parameter changes by transforming the speech library from a static collection of manually recorded segments to a dynamic expanded library that includes synthesized speech generated by the pre-trained model. This changes the composition and size parameters of the speech library, enabling sufficient speech segments for coherent synthesis even when original recorded data is limited.
4Ease of manufacture
If manually-collected original language materials are used to train the speech synthesis model, then the model can be trained with available data, but the speech library remains too small for effective splicing and synthesis
Solution Approach 1:
The patent applies preliminary action by pre-training a speech synthesis model using available speech data before performing the actual speech splicing and synthesis. This pre-trained model is then used to generate synthesized speech that expands the limited speech library, enabling high-quality synthesis without requiring extensive manual data collection at the time of use.
Solution Approach 2:
The patent uses copying by generating synthesized speech copies from the pre-trained model and adding them to the speech library as synthesized language materials. These synthesized speech segments serve as virtual copies that expand the effective size of the speech library without requiring additional physical recording sessions.
Data Source
AI summary
The present disclosure provides a method for processing speech splicing and synthesis and apparatus, a computer device and a readable medium. The method comprises: expanding a speech library according to a pre-trained speech synthesis model and an obtained synthesized text; the speech library before the expansion comprises manually-collected original language materials; using the expanded speech library to perform speech splicing and synthesis processing. According to the technical solution of the present embodiment, the speech library is expanded so that the speech library includes sufficient language materials. As such, when speech splicing processing is performed according to the expanded speech library, it is possible to select more speech segments, and thereby improve coherence and naturalness of the effect of speech synthesis so that the speech synthesis effect is very coherent with very good naturalness and can sufficiently satisfy the user's normal use.