Voice Technology

	FAVORITE
	SITEMAP

home ＞ Voice Technology

Voice Technology

Semantic Recognition

Morphological Analysis

Product Development Release Information

Voice Recognition in VOICE CAIWA being developed by PtoPA consists of two main features: ①　Auto formation of Language Model and ②　Prosody Analyzer. They will be discussed below.

1)Auto formation of Language Model

“Language Model Formation” in voice recognition technology is placed in the hands of expert engineers and is a big barrier to widespread use of voice recognition systems. In such a given situation, VOICE CAIWA adds language model auto formation features anew, and provides an environment where people in general are able to create language models.

Language models in general are created based on a corpus of voice recognition domain. The language model of VOICE CAIWA, however, is automatically formed based on the CAIWA knowledge base. When general users create the CAIWA knowledge base, language models are automatically created that would cover “utterances related to the knowledge base”, and the utterances are entered into the CAIWA knowledge base and make adjustment so they will become voice recognition objects. In other words, VOICE CAIWA can be called a system where user-customized voice recognition features are installed.

Such a feature as “the Auto formation of language model customized by users” can be said to be the greatest feature of our voice recognition technology.

2)Prosody Analyzer

The necessary cognition rate of voice recognition in an ideal sound environment can normally be achieved by adjusting language models and sound models. However the cognition rate is frequently degraded by various elements in normal world environments. For example, unintended recognition results are formed caused by casual noises, faltering tonal voices, or discontinued utterances, or the cognition rate can be greatly lowered on account of extreme volumes (loud/small) of speakers voices. In addition, if the volume control is not correctly adjusted, unnecessary noises can be picked up, resulting in low cognition rate. This casts a serious burden on Voice Recognition, namely to realize robust voice recognition by avoiding undesirable influences to sound environments as much as possible. This is true of voice recognition in our VOICE CAIWA. It is necessary to provide the most robust voice recognition and input the recognition results to CAIWA. The filter of the Prosody Analyzer discerns voice data for voice recognition from other sound date (such as sudden noises or fillers). The Voice Control focuses on the sound volume of voice data, and provides such processes as giving off ‘loud voice warning’ when loud voice is sensed, and ‘low voice warning’ for small voice utterances. In addition, a Voice Volume Controlling feature has been developed to respond to environmental noises more flexibly, and volume control can be performed according to the sound level of normal noises.

Prosody Analyzer is being developed in order for our VOICE CAIWA to realize such “voice recognition with high robust ability” Wave processing technology is introduced into prosody analyzer to realize ‘filter processing’, ‘voice control’ and ‘volume control’.

These features mounted in the Prosody Analyzer allow target voice recognition of proper voice data alone, which will help realize robust voice recognition. This realization of robust voice recognition technology can lead to a ‘language model auto formation feature’ as well as a ‘characteristic of voice recognition technology’ of PtoPA.