High performance digit recognition in real car environments

In this paper, we consider the problem of robust digit recognition in real car environments. We choose to utilize newlycollected CU-Move database [2]. We address the problem using two integrated approaches . First, we consider array processing, enhancement and noise adaptation techniques as an integrated solution. This approach reduced the word error rate (WER) 38.6% and increased word accuracy (WAC) 47.1%, relative to baseline results. Secondly, we use array processing, enhancement, cepstral mean normalization, vocal tract length normalization and MLLR adaptation as an alternative solution. The net gain obtained with this solution is 55.4% reduction in WER and 64.3% increase in WAC, relative to baseline results. The first approach has the advantage of speed since all operations can be performed in real-time, while the second approach maintains high accuracy at the cost of increased computational requirements.