To have a meaningful result from the Fourier transform the input signal must filtered to a bandwidth at most half of the sample rate.

Your 10khz maximum sample rate which I suppose is given by the maximum sample rate of the ADC is less than double of 6.61 kHz , the base frequency of your signal.

Even more, you limit the signal adding high frequency components to the signal that already doesn't fit into the required bandwidth.

You better stick to zero cross detection using both rising and falling edges to have more data to average.

You might also check which part of the signal has the correct speed data against the real speed or to check for a pattern, I see the middle period is lower then at the begining and the end of the signal.

To have a meaningful result from the Fourier transform the input signal must filtered to a bandwidth at most half of the sample rate.

Your 10khz maximum sample rate which I suppose is given by the maximum sample rate of the ADC is less than double of 6.61 kHz , the base frequency of your signal.

Even more, you limit the signal adding high frequency components to the signal that already doesn't fit into the required bandwidth.

You better stick to zero cross detection using both rising and falling edges to have more data to average.

You might also check which part of the signal has the correct speed data against the real speed or to check for a pattern, I see the middle period is lower then at the beginning and the end of the signal.

Later edit. Some things you should check and explanation on why is pointless to use FFT on this.

First, I see you have a clean middle centered signal between -0.7 and 3V. Maybe there is an output capacitor on the sensor that auto limits itself between Zener voltage and the lower clamp diode. The middle is somewhere at 1.5V. But the middle of the logic voltage is somewhere at half VDD, 2.5V.

Maybe that's why you have variations of pulse length from the beginning to the end of the train pulse because the decision line is above the middle line.

Of course setting the decision around the middle line might give a lot of noise but you can either use an envelope detector to see when the pulse begins or a Schmidt trigger so only signal above the hysteresis will pass.

I think that truly detecting the passes around the middle point will give you more reliable and consistent results.

About using discrete Fourier transform (fast or not) for finding the base frequency of the signal, there is no way to obtain a reasonable resolution using such a short signal.

The DFT is discreete not only in time but in frequency to. Anybody who has some basic knowledge about DFT knows that the maximum resolution in frequency that you can obtain in a t timeframe is 1/t, here for 2.5ms the frequency will be in steps of 400Hz, if you have maximum 8kHz signal you will have 20 fixed frequency values from 0 to 8kHz no matter how many samples you take, no matter on what software you make the processing or platform. It's a dead end.