Testing and tuning of speech recognition systems using synthetic inputs

Abstract

A system and method of testing a speech recognition system by providing pronunciations to the speech recognizer. First a text document is provided to the system and converted into a sequence of phonemes representative of the words in the text. The phonemes are then converted to model units, such as Hidden Markov Models. From the models a probability is obtained for each model or state, and feature vectors are determined. The feature vector matching the most probable vector for each state is selected for each model. These ideal feature vectors are provided to the speech recognizer, and processed. The end result is compared with the original text, and modifications to the system can be made based on the output text.

Description

technical field [0001] The present invention relates to speech recognition, and more particularly to testing and tuning of speech recognizers. Background technique [0002] First, a basic description of processing used in the speech recognition system will be described. In a speech recognition system, an input speech signal is converted into words representing the spoken content of the speech signal. The conversion begins by converting the analog voice signal into a series of digital values. The digital values ​​then pass through a feature extraction unit which computes a sequence of feature vectors based on the digital values. Each feature vector is typically multidimensional and represents a single speech signal frame. [0003] To identify the most likely sequence of words, apply the feature vector to one or more models that have been trained using the training text. Typically, this involves applying feature vectors to a frame-based acoustic model, where a single frame...

AI Technical Summary

Problems solved by technology

Errors introduced by language models are situations where a speech recognition system is able to recognize individual words when they stand alone, but not in the context in which those words exist in the test

For example, if the language model can recognize "tohose" alone, but not "want to hose" (for example, the system can recognize the input as "want to host"), this is a language model error

In the second example of this error, the language model will correctly recognize "July25th", but not "July 25th."

Images