A Robust Low-Complexity Algorithm for Voice Command Recognition in Adverse Acoustic Environments

摘要

The paper proposes a noise robust, low-complexity algorithm for voice command recognition. The matching block technique, called Vector Quantization-Weighted Hit Rate (VQWHR), uses a time normalization like dynamic time warping (DTW) with the advantage of about 35% of the complexity and about 10% of the memory requirements of the reference technique. Results obtained in tests are given in terms of recognition rate using the TIMIT-46 database with various types of background noise and different SNRs. The experimental results and comparisons with the main matching techniques demonstrate that, offering a good trade-off between performance and complexity, the new algorithm is appropriate for all applications that present great limitations in terms of processing capacity and memory. Finally, for application where it is necessary to recognise linked lists of words, such as names and surnames in voice dialling applications, a contextual analysis based on double matching control is proposed. Thanks to the double matching linking mechanism proposed in this paper, the performance is better than in recognition of names or surnames alone.