Can neural network be able to convert analog timeseries to discrete timeseries

Question

I have timeseries signal of sound echo back from instruments. They are firing kick, marker, wall, echo noise.

Since neural network can produce Seq2Seq prediction. I am using ResNet + 1DCNN to produce the predicted sequence.

What I have tried so far...

First picture blue is prediction, orange(lower line) is solution. The solution is describe by number as follow

class Label(Enum):
    BLANK: int = 0
    LIQUID_LEVEL: int = 1
    ECHO: int = 2
    DOWN_HOLE_MARKER: int = 3
    FIRING_KICK: int = 4

The first picture is made by using only 1 data and training. I found that neural network be able to generate as discrete timeseries.

source file:

However, when I train with more complex network and I give more than 1 data. A.I. started to not be able to generate sharp answer.

Question: Can neural network be able to convert analog timeseries to discrete timeseries?