Autoencoder Text-to-Speech Synthesis Reducing Computational Load

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional text-to-speech synthesis methods face challenges in efficiently generating natural-sounding speech due to high computational requirements and limited ability to represent variable-length audio units effectively, leading to suboptimal quality and increased latency.

Innovation Solution

A text-to-speech system utilizing an autoencoder network with a linguistic encoder and an acoustic encoder, both incorporating long short-term memory layers, to generate fixed-size speech unit representations that can be used for unit-selection synthesis, reducing computational complexity and improving output quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If conventional deep neural network based speech synthesis is used with frame-level acoustic representation, then speech quality can be improved, but computational requirements increase significantly

Engineering Contradiction:
Improvespeech qualityVSAvoidcomputational requirements
Core Design Contradiction:
Manufacturing precisionVSPower

Solution Approach 1:

The patent extracts the essential acoustic characteristics of speech units and represents them using fixed-dimensional bottleneck features. By taking out only the most important acoustic properties and encoding them in a compact fixed-size representation, the system maintains speech quality while significantly reducing computational requirements compared to full frame-level representations.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter representation from variable-length frame-level acoustic features to fixed-dimensional bottleneck features. This parameter transformation allows the system to capture essential speech characteristics in a compact form, reducing computational complexity while preserving speech quality through the use of autoencoder-based dimensionality reduction.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If variable-duration speech units are mapped to fixed-dimensional representations, then computational efficiency is improved, but the ability to represent acoustic characteristics accurately may be degraded

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidacoustic representation accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent introduces an autoencoder-based bottleneck representation as an intermediary between variable-duration speech units and fixed-dimensional vectors. This intermediary structure learns to compress variable-length acoustic sequences into fixed-size representations that preserve essential speech characteristics, enabling both computational efficiency and accurate acoustic representation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms variable-duration speech units into fixed-dimensional representations by projecting them into a different dimensional space using autoencoder bottlenecks. This dimensional transformation allows the system to handle speech units of varying lengths uniformly while maintaining acoustic fidelity through the learned compression in the bottleneck layer.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If text features are repeated for the entire duration of phoneme mapping, then mapping accuracy is improved, but processing time increases

Engineering Contradiction:
Improvemapping accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential linguistic and acoustic features needed for speech synthesis and represents them in fixed-dimensional form. By extracting and representing only the critical features rather than repeating full text features throughout phoneme duration, the system maintains mapping accuracy while significantly reducing processing time.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3376497B1Text-to-speech synthesis using an autoencoder
Publication Date: 2023.12.06 GOOGLE LLC
  • EP3376497B1 patent drawingFigure 1A
  • EP3376497B1 patent drawingFigure 1B
  • EP3376497B1 patent drawingFigure 2

AI summary

Methods, systems, and computer-readable media for text-to-speech synthesis using an autoencoder. In some implementations, data indicating a text for text-to-speech synthesis is obtained. Data indicating a linguistic unit of the text is provided as input to an encoder. The encoder is configured to output speech unit representations indicative of acoustic characteristics based on linguistic information. A speech unit representation that the encoder outputs is received. A speech unit is selected to represent the linguistic unit, the speech unit being selected from among a collection of speech units based on the speech unit representation output by the encoder. Audio data for a synthesized utterance of the text that includes the selected speech unit is provided.