Katsuyuki Niyada

Estimation of Active Speaker’s Direction Using Particle Filters for In-Vehicle Environment

Building in-vehicle human-machine interfaces, information on speaker’s direction is helpful for achieving robust speech recognition and video camera control applications. This chapter presents two-step particle filtering approach in a spectro-spatial domain for achieving robust Direction-Of-Arrival (DOA) estimation in noisy environments such as in-vehicle conditions. The particle filter is applied to track the movement of an active speaker while speaking, and the two-step filtering aims at combining the advantages of both traditional cross-correlation (CC) and generalized cross-correlation (GCC) methods.

Directions of arrival (DOA) estimation by particle filtering with dominant frequency components in noisy environments

In noisy environments, DOA estimation is useful for achieving attractive human‐machine interfaces in the real world. This article proposes a robust DOA finder based on cross correlations between observed signals by particle filtering with frequency selectivity. The particle filter is flexible for modeling the system because it requires no linearity or Gaussianity on the model. The temporal trajectory of DOAs is modeled using a Markovian state space model. It is estimated through the state estimation procedure using particle filters. Markov modeling is expected to achieve stable DOA estimation in time. It is also a key for the proposed method to selectively use the dominant frequency components, such as formant frequencies of speech, to improve noise robustness. Assuming the quasi‐stationarity of noise signals in the presence of a time‐varying target speech signal, the dominant frequencies are extracted using a deterministic approach similarly to spectral subtraction. A set of the subband cross‐correlation...