Phrase-Based End-to-End Text-to-Speech Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional end-to-end text-to-speech (TTS) systems face challenges in speed and stability of speech synthesis due to their autoregressive nature, which affects online service provision and style consistency.
Innovation Solution
The proposed method employs phrase-based end-to-end TTS synthesis, where speech is synthesized in parallel at a phrase level, considering both phrase context and reference audio to control prosody and style effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional end-to-end TTS systems use autoregressive generation, then speech quality can be maintained, but synthesis speed becomes slow and efficiency is reduced
Solution Approach 1:
The patent segments the input text into multiple phrases and generates speech for each phrase in parallel rather than sequentially. This segmentation enables the system to process different phrases simultaneously, significantly improving synthesis speed while maintaining quality through context-aware generation for each phrase segment.
Solution Approach 2:
The patent transitions from sequential single-phrase generation to parallel multi-phrase generation by adding the dimension of concurrent processing. By organizing generation across multiple phrases simultaneously and using context embeddings to maintain coherence, the system achieves faster synthesis without sacrificing quality.
2Productivity
If phrase-based parallel synthesis is implemented, then synthesis efficiency improves, but system complexity increases due to context management requirements
Solution Approach 1:
The patent introduces context embeddings as intermediary representations that capture the semantic and prosodic context of each phrase. These context embeddings serve as mediators between the parallel phrase generation processes, enabling the system to manage complexity by providing a structured way to handle contextual information without requiring complex inter-process communication.
Solution Approach 2:
The patent transforms the complex problem of maintaining style consistency across parallel phrases into a parameter-based solution by using context embeddings with specific dimensions (e.g., style tokens, prosody parameters). This parameterization simplifies the control mechanism, allowing the system to manage multiple parallel generations with standardized contextual parameters rather than complex rule-based systems.
3Reliability
If autoregressive TTS is used, then style consistency can be maintained, but the system lacks effectiveness for online service provision
Solution Approach 1:
By segmenting text into phrases and processing them in parallel with context awareness, the system achieves both fast response times for online services and style consistency through shared context embeddings. Each phrase is generated independently but with reference to the overall context, enabling real-time performance while maintaining reliability.
Solution Approach 2:
The system performs preliminary processing by creating context embeddings that capture style and prosodic information before the actual speech generation. This preliminary action prepares the contextual framework in advance, allowing parallel phrase generation to proceed quickly while maintaining style consistency, thus enabling effective online service provision.
Data Source
AI summary
The present disclosure provides methods and apparatuses for phrase-based end-to-end text-to-speech (TTS) synthesis. A text may be obtained. A target phrase in the text may be identified. A phrase context of the target phrase may be determined. An acoustic feature corresponding to the target phrase may be generated based at least on the target phrase and the phrase context. A speech waveform corresponding to the target phrase may be generated based on the acoustic feature.