| Simmortel Voice Vocalised Services! | | Print | |
| Start Up | |
|
Page 1 of 3 Using voice recognition technology to cater to customers speaking regional languages was, until now, an unexplored concept. Simmortel Voice offers a pioneering service. From interactive voice response technology, the world has moved to voice recognition technology. Earlier, the user pressed the necessary phone keys and the relevant computerised voice response track played. The system did not require or register a user's speech. Voice recognition technology, however, picks up spoken keywords, which the computer can match, enabling it to access the requested information accordingly. Simmortel Voice has launched a platform that incorporates technologies that can recognise a caller's language and accent. The company has developed a voice recognition engine, and voice recognition language models for Indian languages and accents. The engine and models developed by Simmortel Voice are meant to be used in mobile phone and telephone applications. The voice is carried over to the computer server connected to the telephone. Simmortel Voice hosts the server. It converts the voice signal to words, and then performs some action like database look-up, and plays back a response. The technology Voice recognition technology that converts voice signals into words using digital signal processing like Fourier analysis and statistical modelling, backs this platform. Another important aspect of this platform is the telecom grade server. It is a highly available server for processing millions of phone calls in a multi-tenant architecture using commodity class servers. How does it work for Indian languages? There are two aspects to speech: language and acoustics. The acoustics is studied with the help of Fourier analysis, statistical signal processing and hidden Markov models. The language part is all about pronunciation and phonetics. Given a signal in the time domain, it is converted into its frequency domain representation using Fourier analysis and related transforms. The reason for transforming the signal into the frequency domain is for delineation of the distinguishing speech features. Information about pitch and intonation, which are speaker dependent, is filtered out while features rich in acoustics are retained. Thereafter, the features are further transformed to neutralise any remaining effects of pitch. However, just having these features is not enough to apply a simple pattern-matching technique, because speech is a stochastic signal. Speech features vary a lot for the same speaker and the same speech. Therefore, a statistical model is built using these features from training data. This model is called the hidden Markov model, using which, we can predict if a given signal belongs to a particular model. To make it work for Indian languages, one has to train the process for Indian languages. Acoustic models are built using the phonetic knowledge of Indian languages, and language models are built using linguistic and statistical knowledge. Experts from linguistics, phonetics, acoustics, statistics, computer science and electrical engineering collaborate to make this entire process work. "We study different ways in which people pronounce the same words; also, different phones are used. All this knowledge goes into making more robust acoustic models. This effort has to be repeated for each new language. That said, we are still not done with applications for dictation, which is truly language dependent. This is still in the R&D stage. Applications that we are focusing on are actually not tied to a particular language. They are tied to more than one language. For example, names of trains, stations, people, places, businesses, movies, songs, etc, in India belong to many languages. We are making such applications possible," confirms Abhishek Singh, CEO and co-founder, Simmortel Voice. "What we are doing has been done for American English around eight years ago. However, doing everything from scratch for new languages is a technological barrier, which gives us a solid foundation for our business," he adds. |
|
Magazine Archives
| February 2009 |
| January 2009 |
| December 2008 |
| November 2008 |
| October 2008 |