A comparison and combination of segmental and fixed-frame signal representations in NMF-based word recognition

Segmental and fixed-frame signal representations were compared in different noise conditions in a weakly supervised word recognition task using a non-negative matrix factorization (NMF) framework. The experiments show that fixed-frame windowing results in better recognition rates with clean signals. When noise is introduced to the system, robustness of segmental signal representations becomes useful, decreasing the overall word error rate. It is shown that a combination of fixed-frame and segmental representations yields the best recognition rates in different noise conditions. An entropy based method for dynamically adjusting the weight between representations is also introduced, leading to near-optimal weighting and therefore enhanced recognition rates in varying SNR conditions.