librosa load with changed sr and ffmpeg have different result

Question

I have input.wav with sample rate 4800.

1. librosa load in python

import librosa
y, sr = librosa.load("input.wav", mono=True, sr=16000)
print(y)

2. ffmpeg transform and load

% ffmpeg -i input.wav -y -ar 16000 -ac 1 output.wav 

import librosa
y, sr = librosa.load("output.wav", mono=True, sr=16000)
print(y)

I expect first and second have same result. But they are different. How can I fix it?