Efficient blind estimation of subband reverberation time from speech in non-diffuse environments

To respond to reverberation effectively, modern speech processing tasks like signal enhancement in digital hearing aids or distant-talking speech recognition often require precise knowledge of the acoustic situation. A common measure is the (frequency-dependent) reverberation time T₆₀. Explicit measurement using sound excitations is not possible for most applications, therefore blind estimation of subband T₆₀ from speech signals is employed. Previous approaches are either limited to T₆₀ between 0...1.5 s, or require long speech signals and costly computation. We propose an efficient algorithm that estimates T₆₀ up to 6 s from short speech signals of no more than 20 s. In experiments with real room impulse responses (RIR) the algorithm exhibits state-of-the-art performance for T₆₀ up to 1.5 s and superior performance for longer T₆₀.