Changing how long an audio file runs is not something a filter can do. Every effect in a signal chain hands back the same number of samples it was given, so speed is applied after the chain, on the way into the encoder.
Tempo without transposition
With Keep the original pitch on, two things happen in order. A phase vocoder shifts every partial against the speed change, then the stream is resampled to its new length. A 1.5x export is a 7.02 semitone shift down followed by a 1.5x resample, and the two cancel in frequency while both shorten the file.
The window is 46 ms. That is long enough to resolve a bass note and short enough to keep speech intelligible, and it is what sets the character of the artifacts: sustained material survives large changes, percussive material does not.
The clean range and the far end
Between 0.5x and 2x the correction stays within an octave and the result reads as the same recording, retimed. This is the range for lectures, interviews, transcription, and learning a part off a record.
Past an octave the reconstruction outgrows the window. Drum hits spread across it, and a metallic ring builds under sustained notes. At 0.1x that ring is the whole point: an eight or ten times slowdown with pitch held is how a three minute track becomes half an hour of drifting pads, and the smearing is what makes it sound that way.
Speeding past 2x with pitch held is the least useful corner of the range. Speech stops being intelligible before the artifacts matter.
Varispeed
Turning pitch correction off leaves a plain resample. Frequency and duration move together, which is what a tape machine or a turntable does at the wrong speed. A 2x export doubles every frequency, an octave up; 0.1x drops it by more than three.
This is the honest mode for sound design and for checking whether a sample was recorded at the wrong rate. It is also cleaner than the corrected mode, because nothing is being reconstructed.
Anti-aliasing on the way up
Speeding a file up moves content upward in frequency, and anything that lands above half the sample rate folds back down as a tone unrelated to the music. A Butterworth low-pass cuts above the new limit before the interpolator runs.
The order of that filter rises with the speed, because the band being discarded grows with it. 2x throws away the top half of the spectrum, and 10x all but a tenth of it, so a slope that is enough for the first is one you can hear through at the second.
Slowing down cannot alias, so the filter is not built at all in that direction.
Preview against the export
The transport plays the real file at the rate you set, and the pitch correction you hear is the same step the export runs. Far from 1x, browsers mute the player rather than play it, so the preview goes quiet while the export keeps working.
Practical speeds
1.25x and 1.5x are the reading speeds most people settle on for a lecture or an interview. 0.75x is slow enough to transcribe from without words running together. Musicians learning a part usually want 0.6x to 0.8x with pitch kept, so the fingering still matches the recording.