To improve the posterior belief receiving a time-series from a fixed data source Let data $\mathbf{X}\in \mathbb{R}^d$ come from one of the $K$ possible sources $\mathsf{S}\in \{1,2,...,K\}$. The true $\mathsf{S}$ is unknown but it is fixed. The main task is to infer the true value of $\mathsf{S}$ by observing a sequence of $\mathbf{X}$s. We have an uninformative prior on the true $\mathsf{S}$ (i.e. uniform). 
We also have access to an emprical classifier $\mathcal{F}: \mathbf{X} \rightarrow [0,1]^K$ that produce a softmax output reciving $\mathbf{X}$.
As an example, we are observing an audio signal stream and we want to infer the age of the speaker which is something in range $\mathsf{S}\in \{18,19,...,60\}$. We already have an age classifier, $\mathcal{F}$, that has been trained on a public dataset (not including the current speaker's data).
Or, let's say:


*

*We receive $\mathbf{X}_1$ (at time $t=1$) and our previously trained classifier produces a guess about $\mathsf{S}_1$.

*Then, we receive $\mathbf{X}_2$, and we again use the classifier to guess $\mathsf{S}_2$.

*Then, we receive $\mathbf{X}_3$, and so on.


So, at time $t=T$ we have $\mathsf{S}_1, \mathsf{S}_2, ..., \mathsf{S}_T$. 
Now, the questions: 


*

*By having multiple consecutive observations $\mathbf{X}_1, \mathbf{X}_2, \mathbf{X}_3,...$ (e.g. second by second audio signals), how can I improve my belief about the true $\mathsf{S}$ (e.g. speaker's age) over time, leveraging the available $\mathcal{F}$ (e.g. audio based age classifier): 

*Is there any framework to calcualte some sort of confidnece interval over the predicted value and narrowing it down by obseving more and more data?


*

*A reference to a related subject or a link to the state-of-the-art method/article is also appreciated.
 A: Question 1 : Your model gives you $\mathbf{X}_i = \mathbb{P}(S = s | data[t_i:t_{i+1}])$. I think you want $\mathbb{P}(S = s | data[1:T])$. According to the Bayes rule : 
$$\mathbb{P}(S = s | data[1:T]) = \mathbb{P}(data[1:T]| S = s)\mathbb{P}(S = s)/\mathbb{P}(data[1:T]).$$
We have that $data[1:T] = \cup_{i=1}^N data[t_i:t_{i+1}]$. We may want to assume that $data[t_i:t_{i+1}]$ are independent and independent conditionally on S (*). 
$$\mathbb{P}(data[1:T]) = \prod_{i=1}^N \mathbb{P}(data[t_i:t_{i+1}])$$
$$\mathbb{P}(data[1:T]| S = s) = \prod_{i=1}^N \mathbb{P}(data[t_i:t_{i+1}]| S = s).$$
Therefore, we have 
$$\mathbb{P}(S = s | data[1:T]) = \mathbb{P}(S = s)\prod_{i=1}^N \mathbb{P}(data[t_i:t_{i+1}]| S = s)/ \mathbb{P}(data[t_i:t_{i+1}]).$$
We use again the Bayes rule : 
$$\mathbb{P}(S = s | data[1:T]) = \mathbb{P}(S = s)\prod_{i=1}^N \mathbb{P}(S = s|data[t_i:t_{i+1}])/ \mathbb{P}(S=s) = \frac{\prod_{i=1}^N \mathbf{X}_i(s)}{\mathbb{P}(S=s)^{N-1}}.$$
Now we add your assumption on uniform prior on S : $\mathbb{P}(S = s)=1/K$
$$\mathbb{P}(S = s | data[1:T]) = K^{N-1} \prod_{i=1}^N \mathbf{X}_i(s).$$
This is weird because it is not well normalized... Maybe I did some mistakes. But anyway, what I would do is to compute : $\hat{p}(s) = \prod_{i=1}^N \mathbf{X}_i(s)$, and then normalize on $s$ to have $\mathbb{P}(S = s | data[1:T])$. 
Question 2 : Once you have $\mathbb{P}(S = s | data[1:T])$, you can estimate the probability of any set of values $\mathcal{S}$ by summing $\sum_{s \in \mathcal{S}}\mathbb{P}(S = s | data[1:T])$. In particular, you can have the probability of a range of value around the mean, ie a confidence interval.
Of course, your result will be super-dependent on the quality of your model $\mathcal{F}$. For instance, if your model output some inacurrate $0$ in the $\mathbf{X}_i$ values, it will totally screw up your predictor. This should not happen in a correctly train model but you may want to replace extremely low value by a $\epsilon$ "just in case". 
(*): These assumptions may be inaccurate. A more correct way to proceed is to retrain $\mathcal{F}$ to have a sequential model that gives directly $\mathbb{P}(S =s | data[1:T])$ instead of $\mathbb{P}(S =s | data[t_i:t_{i+1}])$ (like LSTM or this kind of stuff). If this is not possible, I guess you need assumptions of this kind to compute more or less meaningful probabilities.
