Hierarchical Prosody Encoder for Natural Text-to-Speech

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing text-to-speech (TTS) models struggle to accurately model prosody, resulting in synthesized speech that lacks expressiveness and fails to convey intonation, stress, rhythm, and style effectively.

Innovation Solution

An attention-based clockwork hierarchical variational encoder (CHiVE) that uses hierarchical LSTM blocks to predict syllable durations, pitch, and energy contours based on linguistic features, encoding reference audio signals into fixed-length utterance embeddings to generate realistic prosody.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional concatenative and parametric synthesis models are used, then intelligible speech can be generated, but the synthesized speech lacks naturalness and expressiveness

Engineering Contradiction:
ImproveintelligibilityVSAvoidnaturalness
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The patent segments speech synthesis into multiple hierarchical levels: phoneme level for basic sound generation, syllable level for duration and rhythm control, and utterance level for overall prosody management. This segmentation allows each level to specialize in specific aspects of speech production, improving both intelligibility and naturalness simultaneously

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a hierarchical temporal dimension by processing speech at multiple time scales concurrently - fast phoneme-level processing, medium-speed syllable-level processing, and slow utterance-level processing. This multi-dimensional approach enables the system to capture both fine-grained articulation details and broader prosodic patterns

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

2Ease of manufacture

If neural modeling is used to improve naturalness, then speech becomes more natural, but the ability to model prosody accurately deteriorates

Engineering Contradiction:
ImprovenaturalnessVSAvoidprosody modeling accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent applies local quality by assigning different neural network architectures and processing strategies to different hierarchical levels. The phoneme-level network focuses on articulation accuracy, the syllable-level network on duration and rhythm, and the utterance-level network on overall prosodic contours. Each level optimizes for its specific local function rather than attempting to handle all aspects uniformly

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent implements feedback mechanisms where higher-level prosodic predictions influence lower-level generation. The utterance-level prosody predictions provide guidance to syllable-level duration modeling, which in turn constrains phoneme-level timing. This hierarchical feedback ensures consistent prosody throughout the synthesis process

Inventive Principle:
Principle #23Feedback

3Productivity

If simple text input is used, then processing is fast and simple, but the ability to convey context and emotion deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidcontext and emotion
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent performs preliminary action by pre-processing text input to extract linguistic features, semantic information, and contextual cues before the main synthesis process. This preprocessing stage prepares enriched representations that capture emotion and context, which are then used by the hierarchical neural networks to generate expressive speech without significantly increasing overall processing time

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4073786B1Attention-based clockwork hierarchical variational encoder
Publication Date: 2026.03.11 GOOGLE LLC
  • EP4073786B1 patent drawingFigure 1
  • EP4073786B1 patent drawingFigure 2A
  • EP4073786B1 patent drawingFigure 2B

AI summary

A method (400) for representing an intended prosody in synthesized speech includes receiving a text utterance (310) having at least one word (240), and selecting an utterance embedding (204) for the text utterance. Each word in the text utterance has at least one syllable (230) and each syllable has at least one phoneme (220). The utterance embedding represents an intended prosody. For each syllable, using the selected utterance embedding, the method also includes: predicting a duration (238) of tire syllable by decoding a prosodic syllable embedding (232, 234) for the syllable based on attention by an attention mechanism (340) to linguistic features (222) of each phoneme of the syllable and generating a plurality of fixed-length predicted frames (260) based on the predicted duration for the syllable.