even for the world’s biggest technology companies, the speech recognition is a difficult problem. Apple and Google collected a large amount of real speech patterns, to adjust the speech recognition algorithm has Siri and Google Now. And although these tools is great, but they still take a long time to recognize your voice commands.
driven build voice application even harder for smaller companies, because they can’t get so many resources like apple and Google, that is to say, you don’t like those big company collected a large number of real voice commands. “If only one company in research and development, you can’t get a good example.” Alexandre Lebrun said.
this is established for Wit. Ai, this service can help developers to integrate their speech samples collected together to build a to be able to compete with apple and Google voice and natural language recognition system. In the next few years as the next wave of voice interface of the development of science and technology, such as no screen networking devices and wearable devices, the service might play a big role.
Wit. Ai has just started, but it has attracted thousands of developers involved in testing services, on Wednesday when it announced the VCS Anderessen Horowitz the seeds of a $3 million investment.
speech recognition: the elephant in the room
Wit. The establishment of ai from Lebrun VirtuOz disappointment against the company. VirtuOz as AT& T the company’s development of speech recognition system, the problem is, every time to build a new system company team must be from scratch.
for each system, they have to collect a new set of sample to adjust the sound. In many cases, the requirements of different customers there is some overlap between instruction part of the recognition, but the company does not bring a customer’s voice samples in another customer’s system use again.
“no matter how hard we try, the problem was obvious – speech recognition impossible perfection.” He wrote in a blog, “in fact, the end user experience is sometimes disastrous. Worse, because the sound integration into the system to spend big price, will not have a supplier for small companies and a developer’s needs.”
last year, Lebrun sold VirtuOz to help run the Siri’s voice recognition company Nuance, then created Wit. Ai.
Wit. Ai work
in general, speech recognition of developers to create a “model”, is that you want the computer to identify a collection of words and phrases, and then to enter different people said these words as much as possible phrases samples to “train” computers to identify them. Due to different way under different user instructions, the model needs to be perfect as far as possible, to identify different expressions of the same order.
essentially Wit. Ai is want companies to share model and adjust the data, like software developers on making share code.
its business model and making similar, making people are free of any public share encoding, Wit, ai for sharing data is also free of charge. Due to privacy issues and operability, used to adjust the system of the real recording will not be Shared. And those who don’t want to open their own model and the data of the company can pay to use the service.
provide free service for
Wit. Ai joined a growing number of companies and projects, including some open source projects such as Julius and CMU Sphinx and hosting service such as Google voice text conversion, is dedicated to helping developers to join the speech recognition in their application, so you can better identify voice, find out what the user’s requirements.
Lebrun hope by providing free services to attract a large number of models and adjust the data, in order to provide to rival apple and Google’s speech recognition service.
one problem is that all audio through the Internet to the company’s servers, which means that the delay may be produced, availability, and privacy issues. But Lebrun said they are developing a major with the server running on the client and then mixed version of exchange of information.