Speech Recognition Technology for India
February 10, 2005
Press release written by Madelaine Plauche and Leah Hitchcock-Ybarra
February 8, 2005
Researchers at ICSI will demonstrate their speech recognition technology developed for India with UC Berkeley's TIER Project (Technology and Infrastructure for Emerging Regions) as part of the UCB BEARS Conference on Thursday, February 10th.
Members of ICSI's Speech Group are working to provide speech recognition technology to UC Berkeley's TIER project (Technology and Infrastructure for Emerging Regions). Technologies developed for the affluent world and imported to developing regions often fail to address key challenges in cost, deployment, power consumption, and support for semi-literate and illiterate users. This issue prompted Chuck Wooters and Madelaine Plauche to begin developing a speech recognizer for Tamil, a language spoken by over 50 million people in Southeast India, where illiteracy rates hover around 50% for men and between 60% to 80% for women. Speech recognition, especially in combination with text-to-speech and visual user interfaces, may be key in increasing access to technology to those with limited or no literacy.
One of the many challenges these researchers face is how to develop speech recognition that can support different dialects or accents. Tamil, like most languages, refers to several mutually understandable dialects varying by geography, social factors (caste), and register (formal vs. informal). Tamil is itself one of dozens of languages spoken in India, suggesting that many users of technology will be using a language to interact with that technology that may not be their primary language. An ideal speech recognition system would support the small but significant variations in pronunciation due to these factors.
UC Berkeley researchers designed a data collection system for a Tamil speech recognition system, and have collected data on 30 words in Tamil using 8 speakers at the UCB campus and 22 speakers in India. In February, Plauche will travel to three different sites in Tamil Nadu, India to collect data of more native speakers saying digits and command words in Tamil. She hopes to sample the speech of both uneducated and educated speakers, urban and rural speakers, and speakers from three different geographical dialects, to further investigate how dialect and geographic location may affect recognition error rates.
Plauche and Wooters have also created a sample speech recognition application called Tamil Market, a simulated 1-800 number that allows farmers and other rural community members to get information on market prices for agricultural crops, local weather, and agricultural innovations over the telephone. Tamil Market is for information inquiry, relies exclusively on speech recognition for user interface, and runs on a vocabulary of only 30 Tamil words: digits, some crop names, and selected command words. By allowing Tamil speakers in both urban and rural areas in Tamil Nadu to test this application, they hope to learn how speech recognition can be successfully integrated into useful applications for developing regions.
The Tamil Speech Recognizer demonstration will be held at the ICSI, 1947 Center Street, Suite 600 at 2:00 p.m. on Thursday, February 10th. This demonstration is part of the ICSI Open House, in conjunction with UC Berkeley's BEARS Conference.