2022-04-07-gbo-ivectors.md

layout	title	tags	mathjax
post	GBO notes: i-vectors and x-vectors	gbo i-vectors	true

In this note, we will review the two most popular speaker embedding extraction methods, namely i-vectors and x-vectors. But first, it would be useful to quickly recap generative and discriminative models.

Suppose we have some observed variables $X$ and some target variables $Y$. In the case of speaker recognition, $X$ can be an audio sample spoken by a speaker, and $Y$ can be the speaker id (or class index from a given set of speakers). A discriminative model directly models the conditional probability $P(Y\mid X)$. In our example, the discriminative model would output a probability distribution over the set of speakers, given any utterance.

A generative model is a model of the joint probability $P(X,Y) = P(X\mid Y)P(Y)$. It can also be used to perform classification, by using the Bayes rule as follows:

$$ P(Y\mid X) = \frac{P(X,Y)}{P(Y)} = \frac{P(X\mid Y)P(Y)}{\sum_Y P(X\mid Y)P(Y)}.$$

Neural network models trained for classification tasks are discriminative models that predict logits which are then converted into a distribution over the targets using softmax.

Generative embeddings with i-vectors

We basically need a way to characterize the variability in acoustic features resulting from speaker and channel differences. Suppose the utterances are represented by a sequence of feature vectors (usually MFCCs). A naive way may be to compute the average over the whole sequence; but this would also capture things like lexical content, phonetic variation, and so on. So what we actually want to do is to compare the difference between the features of the target utterance compared with those from a general set of utterances, in the hope that this difference would remove any effects arising from phonetic/lexical content.

This is done with the help of Gaussian mixture models (GMMs). First, we pool together a large set of utterances and compute their features. We then assume that all the feature vectors are generated I.I.D. using a GMM with a fixed number of components (usually 2048). We use the EM algorithm to learn the parameters of this GMM, and call it the Universal Background Model (UBM).

Suppose we now concatenate the means of all the UBM components into a giant vector (called a "supervector"), and denote it as $m$. Suppose we adapt the UBM for a target speaker (or utterance) and the concatenated means of this "target model" is denoted by $M$. Then, we can make the assumption of factor analysis and claim that

$$ M = m + Tw, $$

where $T$ is a low-rank total variability matrix, and $w$ is the i-vector (total factors). We can then say the the i-vector $w$ will characterize all the variations in the target utterance arising from channel or speaker chracteristics.

The parameters of the model $T$ and $\Sigma$ are estimated jointly with the latent variable $w$ using the EM algorithm. In the E-step, we fix $T$ (randomly initialized) and $\Sigma$ (initialized from UBM), and compute the $w$ for each utterance. In the M-step, we use the sufficient statistics to update the model parameters.

At inference time, the i-vector of the utterance is computed as the MAP estimate of the feature sequence given the model parameters, i.e.,

$$ \hat{w} = \text{arg} \max_{w} \prod_{c=1}^C \prod_{i=1}^N \mathcal{N}(o_i \mid M_c + T_c w, \Sigma_c) p(w), $$

where $C$ is the number of Gaussian components, and $o_1, \ldots, o_N$ is the sequence of feature vectors.

Discriminative embeddings with x-vectors

While i-vectors are mathematically attractive and useful in the sense that they do not need a labeled training set, they are computationally expensive. Training the UBM model and the parameters of the GMM is time-consuming, and since they model the total variability (i.e., speaker and channel), it is hard to tease apart the speaker-specific information.

Deep neural network based speaker embeddings have become popular as a discriminatively-trained alternative to i-vectors. Given a sequence of feature vectors, we use a neural network to obtain an utterance-level representation, which is then fed into a softmax layer that estimates a distribution over a large set of speakers. The parameters of the network are trained using gradients of a classification-style loss.

Conventionally, the x-vector neural network consists of TDNN (basically 1-D strided CNN) layers are the bottom which operate at the frame-level, followed by a statistics pooling layer. This layer aggregates over the whole input sequence and computes its mean and standard deviation. These utterance-level stats are concatenated together and passed to further hidden layers that operate at the segment level and produce the final representation that is fed to the softmax.

At the time of inference, the output layer is discarded and the segment-level representation immediately preceding it is used as the embedding, also called an x-vector.