NEWS & EVENTS
December 30, 2010
GritTec has published last results of researches on tuning of GritTec's Speaker-ID engine in the mode of voice verification.
GritTec has published last results of researches on tuning of
GritTec's Speaker-ID engine in the mode of voice verification with
compared the previous version. All researches were made on real
records for CELL and VOIP channels.
As well as in the previous case for an estimation accuracy of
GritTec's Speaker-ID engine in mode of voice verification it was
used voices of 25 target speakers (12 - males, 13 - females) for
English language. Each target speaker was trained separately for
CELL and VOIP channel. It was used 50 files for training, 25 files -
for CELL channel and 25 files - for VOIP channel. Each trained file
contained 12 phrases of a random digit numbers (from 0 to 5) with
the common duration ~
The total of files used for verification in CELL and VOIP channels
was 31950, where 30195 files with voices of imposter
speakers, and 1755 files with voices of target speakers. Each
verified file contained 1 phrase of a random digit numbers
(from 0 to 6) with the common duration ~ (4-6) seconds.
Area of the last researches it was reduced to adjustment of building
target and UBMs models. In result at using of last parameters it was
possible to receive more lower values of EER errors in comparison
with the previous version.
Screenshots of DET curves with EER (Equal Error Rate) errors of
verification results for CELL and VOIP channels are shown below.
To know more details about GritTec's Speaker-ID, please visit
home page of this
product or contact us via Online Request
GritTec Laboratory (GritTec Ltd.) specializes on research and
development of algorithms and technologies in the field of speech
and audio processing. GritTec's research is focused on speech
enhancement, speech concealment, voice biometric, speech recognition,
speech synthesis and other speech and audio technologies.