I guess there's nothing much you can do with the current API.
You could try to modify the internal sample size: at line 145 of Music.cpp, you can try:
m_samples.resize(1024);
1024 is not a lot so I don't know if the music will still play smoothly. And it will still not be perfect, you'll still be 1024 * 2 samples before what's played (SFML streams are triple buffered).
I don't know if true real-time is achievable without specialized audio APIs (or even hardware?), to be honest I have never thought about this kind of use case.