Parametric Speech Synthesis Reducing RAM via Frame Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing parametric speech synthesis methods require increasing RAM capacity in direct proportion to the length of synthesized speech, making it impossible to continuously synthesize speech of arbitrary time length on a chip with a small RAM capacity due to the need to save temporary and optimized parameters across multiple frames.

Innovation Solution

A longitudinal processing method where each frame of speech is synthesized by extracting rough values, filtering, global optimization, and then synthesizing, with only the current frame's parameters saved, using static parameters and multi-subband unvoiced and voiced sound excitation to maintain natural speech quality without rigid boundaries.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If transverse processing manner is used in hierarchical operations, then speech synthesis accuracy is improved, but RAM capacity requirement increases in direct proportion to speech time length

Engineering Contradiction:
Improvespeech synthesis accuracyVSAvoidRAM capacity
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent segments the speech synthesis process into frame-by-frame sequential processing units. Each frame is processed independently through the five hierarchical layers, with only necessary previous frame data retained. This segmentation allows the system to maintain high synthesis accuracy while limiting RAM usage to a fixed buffer size rather than requiring storage for all frames simultaneously.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements a nested processing structure where frame-level processing is nested within phone-level processing, which is nested within word-level processing. This hierarchical nesting allows the system to manage memory efficiently by maintaining only the necessary context at each level, with inner levels having access to relevant data from outer levels without requiring all data to be simultaneously available in RAM.

Inventive Principle:
Principle #7Nested doll (Nesting)

2Measurement precision

If temporary parameters of all frames are saved in RAM for reverse recursion, then parameter prediction accuracy is improved, but speech of arbitrary time length cannot be continuously synthesized on chip with small RAM

Engineering Contradiction:
Improveparameter prediction accuracyVSAvoidcontinuous synthesis capability
Core Design Contradiction:
Measurement precisionVSDuration of action of moving object

Solution Approach 1:

The patent performs preliminary forward recursion to generate temporary parameters, then immediately uses these parameters in the subsequent reverse recursion step without storing all of them in RAM. By processing frames sequentially and maintaining only a sliding window of recent frames in memory, the system achieves accurate parameter prediction while enabling continuous synthesis of speech of arbitrary duration on chips with limited RAM capacity.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If parameters of all frames are saved for global variance model optimization, then speech naturalness is improved, but RAM capacity requirement increases with speech length

Engineering Contradiction:
Improvespeech naturalnessVSAvoidRAM capacity
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies partial action by optimizing speech parameters using only a sliding window of recent frames rather than all frames simultaneously. This partial optimization maintains speech naturalness by considering sufficient contextual information while dramatically reducing RAM requirements. The system processes frames sequentially, applying global variance model optimization to a manageable subset of frames at each step.

Inventive Principle:
Principle #16Partial or excessive action

4Measurement precision

If rigid boundary determination is used for voiced and unvoiced sounds, then speech parameter clarity is improved, but tone distortion occurs in synthesized speech

Engineering Contradiction:
Improvespeech parameter clarityVSAvoidtone distortion
Core Design Contradiction:
Measurement precisionVSObject-generated harmful factors

Solution Approach 1:

The patent applies local quality by using different processing strategies for different portions of the speech signal. Voiced and unvoiced frames are determined using traditional methods, but the excitation signals are blended with overlapping windows to create smooth transitions. This allows clear parameter determination locally while avoiding harmful tone distortions at the boundaries between voiced and unvoiced segments.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent beforehand cushions the transitions between voiced and unvoiced sound segments by using overlapping windowing functions. The excitation signals for voiced and unvoiced frames are blended with smooth transitions using window functions that overlap by a certain duration. This prior cushioning prevents abrupt changes that would cause tone distortion while maintaining the clarity of speech parameters.

Inventive Principle:
Principle #11Beforehand cushioning (Prior cushioning)

Data Source

PatentEP2579249B1Parameter speech synthesis method and system
Publication Date: 2018.03.28 GOERTEK INC
  • EP2579249B1 patent drawingFigure 1
  • EP2579249B1 patent drawingFigure 2
  • EP2579249B1 patent drawingFigure 3

AI summary

The present invention provides a parametric speech synthesis method and a parametric speech synthesis system. The method comprises sequentially processing each frame of speech of each phone in a phone sequence of an input text as follows: for a current phone, extracting a corresponding statistic model from a statistic model library and using model parameters of the statistic model that correspond to the current frame of the current phone as rough values of currently predicted speech parameters; according to the rough values and information about a predetermined number of speech frames occurring before the current time point, obtaining smoothed values of the currently predicted speech parameters; according to global mean values and global standard deviation ratios of the speech parameters obtained through statistics, performing global optimization on the smoothed values of the speech parameters to generate necessary speech parameters; and synthesizing the generated speech parameters to obtain a frame of speech synthesized for the current frame of the current phone. With this solution, the capacity of an RAM needed by speech synthesis will not increase with the length of the synthesized speech, and the time length of the synthesized speech is no longer limited by the RAM.