In a UBM-GMM system, each speaker's distribution of feature vectors is modeled as a mixture of Gaussians, forming a Gaussian mixture speaker model (GMSM).
For each speaker, the conditional distribution given a specific Gaussian is a 128-dimensional Gaussian distribution, i.e., p ( x a | k a ) = ∏ c = 1 128 N ( x a c ; μ a, k a c, σ a, k a c ) and p ( x b | k b ) = ∏ c = 1 128 N ( x b c ; μ b, k b c, σ b, k b c ), where k a and k b are two Gaussian indices, and p ( x a c | k a ) and p ( x b c | k b ) are one-dimensional Gaussians.
Other qualities which help to define conversation include the equal distribution of speaker rights; mutual respect among speakers; spontaneity and informality; and a non-businesslike ambience.
Extended from text-dependent verification to text-independent verification, a rational inference could be that considering the sparsity in the real-world application and imbalanced distribution of speaker frames, it is better not to adapt those Gaussian components which contribute least in modeling speaker frames.
The distribution of speaker's features is clearly distinguished here.
This is thanks brand's perception as a leading speaker company and great distribution channels such as Best Buy.
The pitch range is estimated using the onset and offset algorithm considering the distribution of speaker energy in the modulation spectrogram domain.
For each speaker i, the probability distribution P i (LVT i, k)) of VTL LVT i, k) for all neutral speech is calculated, and we choose the one with the highest probability as the estimated vocal tract length.
We analyze the intrinsic sparsity and imbalanced distribution of speaker frames, those redundant Gaussian components, are discarded in the i-vector extraction phase so as to compensate the phonetic variability.
