Thumbnail Image

Publication or External Link






Automatic speaker recognition in uncontrolled environments is a very challenging task due to channel distortions, additive noise and reverberation. To address these issues, this thesis studies probabilistic latent variable models of short-term spectral information that leverage large amounts of data to achieve robustness in challenging conditions.

Current speaker recognition systems represent an entire speech utterance as a single point in a high-dimensional space. This representation is known as "supervector". This thesis starts by analyzing the properties of this representation. A novel visualization procedure of supervectors is presented by which qualitative insight about the information being captured is obtained. We then propose the use of an overcomplete dictionary to explicitly decompose a supervector into a speaker-specific component and an undesired variability component. An algorithm to learn the dictionary from a large collection of data is discussed and analyzed. A subset of the entries of the dictionary is learned to represent speaker-specific information and another subset to represent distortions. After encoding the supervector as a linear combination of the dictionary entries, the undesired variability is removed by discarding the contribution of the distortion components. This paradigm is closely related to the previously proposed paradigm of Joint Factor Analysis modeling of supervectors. We establish a connection between the two approaches and show how our proposed method provides improvements in terms of computation and recognition accuracy.

An alternative way to handle undesired variability in supervector representations is to first project them into a lower dimensional space and then to model them in the reduced subspace. This low-dimensional projection is known as "i-vector". Unfortunately, i-vectors exhibit non-Gaussian behavior, and direct statistical modeling requires the use of heavy-tailed distributions for optimal performance. These approaches lack closed-form solutions, and therefore are hard to analyze. Moreover, they do not scale well to large datasets. Instead of directly modeling i-vectors, we propose to first apply a non-linear transformation and then use a linear-Gaussian model. We present two alternative transformations and show experimentally that the transformed i-vectors can be optimally modeled by a simple linear-Gaussian model (factor analysis). We evaluate our method on a benchmark dataset with a large amount of channel variability and show that the results compare favorably against the competitors. Also, our approach has closed-form solutions and scales gracefully to large datasets.

Finally, a multi-classifier architecture trained on a multicondition fashion is proposed to address the problem of speaker recognition in the presence of additive noise. A large number of experiments are conducted to analyze the proposed architecture and to obtain guidelines for optimal performance in noisy environments. Overall, it is shown that multicondition training of multi-classifier architectures not only produces great robustness in the anticipated conditions, but also generalizes well to unseen conditions.