Dynamic Pronunciation Dictionary Generation for TTS Voice Quality

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The process of building a text-to-speech voice for spoken dialog systems is complex and error-prone due to inaccuracies in automatic speech recognition, requiring extensive manual correction and inefficient collaboration among researchers, leading to degraded voice quality and increased costs.

Innovation Solution

A toolkit is developed to generate a TTS voice by dynamically creating a pronunciation dictionary based on dialect, linguistic context, and speaker characteristics, with features for error tracking, phoneme alignment correction, and efficient testing, allowing multiple researchers to work simultaneously and reducing the time and cost of voice generation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If automatic speech recognition is used to label phonemes, then the labeling process is automated and faster, but the accuracy of phoneme labeling degrades due to inherent ASR error rates

Engineering Contradiction:

Improvelabeling speedVSAvoidphoneme labeling accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The system implements feedback by using the initially labeled phoneme data to train and improve the ASR model iteratively. The labeled data from manual corrections is fed back into the system to retrain the acoustic models, progressively reducing error rates in subsequent automated labeling passes.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The system performs preliminary automated labeling using ASR before manual verification, establishing an initial phoneme database that can be quickly improved through iterative refinement rather than starting from scratch with purely manual labeling.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If traditional manual methods are used to correct ASR errors, then labeling accuracy improves, but the time and cost required increases significantly

Engineering Contradiction:

Improvephoneme labeling accuracyVSAvoidvoice building time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system enables continuous improvement by automating the iterative process where ASR labeling, error identification, manual correction, and model retraining occur in continuous cycles, progressively improving accuracy without linearly increasing time investment.

Inventive Principle:

Principle #20Continuity of useful action

Solution Approach 2:

The system performs self-correction by automatically identifying problematic phoneme labels through confidence scoring and statistical analysis, then autonomously correcting obvious errors before presenting only ambiguous cases for manual review, reducing the burden on human operators.

Inventive Principle:

Principle #25Self-service

3Measurement precision

If a comprehensive pronunciation dictionary is created to improve ASR accuracy, then recognition precision improves, but the complexity of the voice building process increases

Engineering Contradiction:

ImproveASR recognition accuracyVSAvoidvoice building process complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the pronunciation dictionary building process into manageable components: automatic phoneme segmentation from audio, statistical analysis of phoneme variants, contextual relationship identification, and iterative model training, making the complex task tractable and systematic.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system dynamically adjusts parameters such as phoneme labeling confidence thresholds, ASR model configuration, and training data selection based on the specific voice being built and the observed error patterns, optimizing the balance between accuracy and processing efficiency for each project.

Inventive Principle:

Principle #35Parameter changes

4Reliability

If extensive manual testing is performed to identify pronunciation errors, then voice quality improves, but the testing time and resources increase dramatically

Engineering Contradiction:

Improvevoice qualityVSAvoidtesting time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The system performs partial testing by using statistical sampling and confidence scoring to identify and focus testing efforts on the most problematic phoneme labels and pronunciation variants, rather than exhaustively testing every single phoneme occurrence, achieving adequate quality verification with reduced time investment.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS7630898B1System and method for preparing a pronunciation dictionary for a text-to-speech voice

Publication Date: 2009.12.08 CERENCE OPERATING CO

AI summary

Disclosed are various elements of a toolkit used for generating a TTS voice for use in a spoken dialog system. The embodiments in each case may be in the form of the system, a computer-readable medium or a method for generating the TTS voice. One embodiment of the invention relates to a method of generating a database for a TTS voice. The method comprises matching every spoken word associated with a TTS voice database with a smallest set of possible pronunciations for each word. The smallest set is generated by automatically determining a dialect and linguistic context using linguistic rules, empirically determining idiosyncratic speaker characteristics and determining a subject domain. The method further comprises dynamically generating a pronunciation dictionary on a word-by-word basis using the smallest set.