Text Normalization for Speech Synthesis Using CRF and Spell Checker
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech synthesis systems struggle with text messages containing slang terms, abbreviations, and non-standard words, leading to imperfect phonetic approximations and distracting output, as they rely on models that map known words to audio, and are unable to effectively handle the variety and complexity of alternative spellings and word forms found in text messages.
Innovation Solution
A method and system using a conditional random field model to generate non-standard tokens from standard tokens, enabling speech synthesis by selecting operations on input characters to produce output tokens that are stored in association with the standard tokens, and utilizing a database query to identify corresponding standard tokens for non-standard tokens, thereby improving text normalization and speech synthesis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If speech synthesis systems use standard phonetic models to synthesize speech from text, then speech synthesis can be achieved for standard words, but the system produces imperfect phonetic approximations and distracting output when encountering slang terms, abbreviations, and non-standard words
Solution Approach 1:
The system performs text normalization before speech synthesis by proactively identifying and replacing nonstandard tokens (slang, abbreviations, misspellings) with their standard dictionary equivalents. This preliminary processing step ensures that the speech synthesis engine receives only standard words it can reliably synthesize, eliminating the problem of imperfect phonetic approximations while maintaining high adaptability to informal language through the normalization dictionary.
2Productivity
If speech synthesis systems synthesize unknown words using phonetic approximations or letter-by-letter spelling, then speech output is generated for any input text, but the output does not follow the expected flow of normal speech and becomes distracting
Solution Approach 1:
The system normalizes text to standard words before speech synthesis, ensuring that the speech engine receives properly formatted input that will produce natural-sounding output. This prevents the degradation to phonetic approximations or letter-by-letter reading while maintaining the ability to generate speech for any input text.
Solution Approach 2:
The text normalization system acts as an intermediary layer between the input text and the speech synthesis engine. It translates nonstandard language into standard language, allowing the speech engine to operate in its optimal range while the system as a whole handles diverse informal input. This intermediary processing preserves speech naturalness without limiting productivity.
3Adaptability or versatility
If existing dictionaries provide translations for common slang terms and abbreviations, then some nonstandard words can be normalized, but the variety of alternative spellings and constructions is too broad to be accommodated by a standard dictionary
Solution Approach 1:
The normalization dictionary is designed to handle multiple types of nonstandard language variations (slang, abbreviations, misspellings, alternative constructions) through a unified mapping system. Rather than requiring separate handling for each variation type, the dictionary provides universal coverage by mapping diverse nonstandard forms to their standard equivalents, achieving high adaptability without proportionally increasing complexity.
Data Source
AI summary
A method for identification of a standard text token in a dictionary that corresponds to a non-standard token identified in text includes identification of a first standard token that is associated with the non-standard using a predetermined conditional random field (CRF) model and identification of a second standard token that is associated with the non-standard token using a spell checker. The method further includes identification of noisy channel scores using data from the CRF model and the spell checker for the first standard token and the second standard token, respectively. The method further includes presentation of one of the first and second standard tokens having the greatest identified noisy channel score to a user with a user interface device.