NEW YORK--Information Storage Devices Inc. today launched an advanced embedded speech-recognition engine that it said is the first of its type to be accurate enough for consumer applications.
The ISD-SR3000 offers speaker-independent speech recognition with continuous speech and digit input, with up to 99% accuracy. It requires no user training, and works well across dialects and genders, according to the company.
The speaker-independent commands and speaker-defined voicetags are stored in external memory. Typical applications include up to 50 commands and 50 voicetags, stored in 256 Kbytes of memory.
The ISD-SR3000, the first in a new family of chips called Simon, was demonstrated at the SpeechTEK Conference today in New York. Simon is a chip-level solution that includes hardware, firmware and voice user interface.
According to ISD, users will have the same type of voice command over cell phones, voice message machines, and handheld devices as are now avaliable only for powerful PC- or server-based products. The company said the product is also ideal for small, low-power consumer products, such as home appliances, as well as automotive and industrial applications.
"Imagine being able to tell your handheld device to display your calendar, or while driving, to verbally command your cell phone to dial a number. Speech recognition technology not only makes complex consumer products easier to use, but also much safer --- for instance, helping to decrease the number of auto accidents due to cellular phone usage," said Larry Gaddy, senior product marketing manager for speech recognition products at San Jose-based ISD. "The Simon engine will also allow disabled persons to operate products purely with voice, thereby enabling increased productivity, greater personal freedom and improved quality of life."
ISD's speech recognition technology uses a sub-phoneme recognition process. The sampled speech is split into distinct phonetic sounds, then Hidden Markov models are used to hypothesize boundaries between sounds, and to form probabilistic models for each possible combination. This process allows a wide variety of speakers to use the technology, and recognizes continuous speech patterns in real time. The speech recognition function is always active, enabling voice activation with keyword commands as well as a push-to-talk option for battery applications.
Demo units of Simon ICs are now available for beta site partners. Simon chips are expected to begin sampling in the first quarter of 2000, and will be priced under $5 in consumer-product quantities.