Cutting-Edge R&D, Lean Development and Large-Scale Industry Know-How
Speech Engineering was founded to exploit some new algorithms in machine listening, specifically focusing on identifying continuous features of speech, such as regional accents and speech impediments, rather than the identification of individual words and phrases.
The founding team all have backgrounds combining innovation with solving the problems of large scale operations in the real world. We have applied this not only to our original algorithms, but also to the deployment of more traditional speech technology, for instance, augmenting systems with modules to harvest new vocabulary from the web, and to distinguish between new words, and anomalous words, from other languages.
The Real World
Even the largest corporations in the speech technology sector release products which are not really ready to leave the laboratory. Hardly a week goes by without new articles being published, explaining how public-facing speech interfaces, fail for large sections of the population, who speak with accents very different from those used for training the systems.
We have encountered large broadcasters who have deployed systems for indexing AV archives using speech recognition, yet have not included any mechanism for adding new vocabulary, so their systems cannot recognise terms like "Brexit", "fracking" or "trumponomics: newly significant names like "Dvokovic", "Rihanna" or "Sanaa", let alone the creative, made-up names of emerging companies, like "Juicero", "Grailbio" or "GoJuno".
As well as democratising the application of speech recognition to the wider population, accent recognition can be applied in fields as diverse as surveillance and education - taking thousands of hours of recorded speech and filtering it using knowledge of suspects, to a manageable - listenable - amount, or assessing a language learner's pronunciation in real-time and training them to speak with a natural, native accent.