Playing a video at high speed without squeaking voices
I always wanted to solve this mystery, but only now I bothered to inquire:
I noticed that, when I play a video or audio file at 110% speed, the voices become high-pitched, even at only 10% overspeed. But when I tell YouTube to play a video at 125% speed, the voices only change ever so slightly. How is YouTube doing that? Is it skipping frames to obtain the extra speed?
❶ Qwen4-Max:
Short answer
No — nothing is being skipped on the audio side, and YouTube isn’t doing anything clever itself. Your player and your browser are performing two mathematically different operations that both happen to be labelled “speed”:
Operation What it does to the signal Duration Pitch Resampling / varispeed plays the same samples at a faster clock rate (like fast-forwarding tape) ÷ r × r Time-scale modification (TSM) rebuilds a shorter signal out of the same material ÷ r unchanged Your local player is doing the first. Browsers do the second by default: the HTML spec defines
preservesPitch, whose default value istrue, soplaybackRate = 1.25is explicitly required to keep pitch. YouTube just setsvideo.playbackRate; the browser’s audio pipeline does the DSP.Why 110% sounds so badly wrong in your player
Resampling multiplies every frequency by r, not just the fundamental:
- 1.10× → +1.65 semitones; F0 120 Hz → 132 Hz, F1 700 Hz → 770 Hz
- 1.25× → +3.86 semitones (almost a major third); F0 120 Hz → 150 Hz, F1 700 Hz → 875 Hz
- 2× → exactly one octave
The reason a mere 10% sounds disproportionate is that formants move too. A real speaker raising their pitch by 1.65 semitones keeps their vocal tract the same length, so the formant pattern stays put. Resampling shortens the apparent vocal tract by 10%, which is a speaker-identity cue, not a prosody cue — your ear hears “different, smaller person”, i.e. chipmunk. At 1.25× it would be grotesque, which is exactly what you’d get on YouTube if you flipped the flag.
Try it: open the console on a YouTube video and run
document.querySelector('video').preservesPitch = false;At 1.25× you’ll instantly get the effect your local player produces.
What the browser actually does: WSOLA
Chromium’s audio renderer contains two algorithms and picks between them: “resampling, which shifts the audio’s pitch, and WSOLA, which doesn’t”. Firefox uses SoundTouch, and Android’s ExoPlayer/Media3 uses the Sonic library for the same purpose. FFmpeg’s
atempofilter is another implementation of the same family.WSOLA = Waveform Similarity Overlap-Add. The intuition:
- Cut the input into overlapping analysis windows of a few tens of milliseconds, tapered with a Hann window.
- Emit them at an output hop of half a window, but advance the input pointer by
tempo × half-window. At 1.25× you therefore consume 25% more material than you emit — you discard 20% of the recording.- The crucial step: before cross-fading the next window onto the previous one, slide it back and forth by up to ±half a window and pick the offset that maximises cross-correlation with what’s already been output. On voiced speech, that optimum lands on an integer number of glottal periods, so the pulses line up instead of colliding out of phase.
The net effect is that you delete roughly one glottal cycle out of every five at 1.25×, splicing at phase-matched points. Each surviving cycle still has its original duration → F0 unchanged; the waveform’s fine structure is untouched → formants unchanged. What changes is only how many cycles a vowel lasts, i.e. segment duration and speaking rate.
So your intuition isn’t entirely wrong: material is being thrown away. But it’s discarded in tiny, pitch-period-aligned, cross-faded slices — not by squeezing the time axis, and not by dropping frames.
(The other main family is the phase vocoder: STFT, rescale the hop, fix up phase relationships, resynthesise. Better on polyphonic music, worse on transients; that’s what Rubber Band and Audacity’s high-quality stretching use.)
Why you still hear something at 125%
It isn’t a residual pitch shift; it’s the cost of the splicing:
- occasional mis-aligned splices → momentary warble/”phasiness”, a slightly metallic or reverberant quality;
- smeared plosive bursts and fricative onsets, since noise-like aperiodic segments have no period to lock onto;
- genuine linguistic change: faster articulation, compressed pauses, altered rhythm — which reads as “hurried” even with identical F0.
There’s a lovely historical footnote here: before 2020, Chromium deliberately used resampling in the range 0.95–1.06, on the grounds that WSOLA’s warbling near 1.0× was worse than the tiny pitch shift. Adding
preservesPitchremoved that compromise.The video side — that’s where frames get dropped
Yes, for picture it’s essentially frame skipping: audio is the master clock, each frame’s presentation deadline is divided by r, and the decoder/compositor drops frames it can’t show in time (and can skip non-reference frames when decoding). That’s harmless perceptually. Audio can’t work that way — dropping every fifth 20 ms buffer would give you 25% of the speech missing in audible chunks, and skipping samples is precisely what raises pitch.
Fixing your player
- VLC: Tools → Preferences → Audio → Enable time-stretching audio must be checked; unchecked, you get varispeed.
- mpv:
--audio-pitch-correctionis on by default and auto-insertsscaletempo2at non-unit speed; for better quality use--af=rubberband.- MPC-HC/BE: use the internal audio renderer; bitstreaming/exclusive-mode passthrough paths bypass pitch correction entirely.
- FFmpeg:
-af atempo=1.1preserves pitch;asetrate=48000*1.1,aresample=48000is the chipmunk version.
I am a fan of MPC-HC:
MPC-HC has such a complex preferences set that I couldn’t find any such setting (version 2.8.1).
❷ Qwen4-Max:
You weren’t missing it — there is no “preserve pitch” checkbox anywhere in MPC-HC. Pitch correction is a property of the audio renderer you select, not a separate option. That’s why hunting through the preferences tree turns up nothing.
The 2.8.1 documentation states it explicitly: ”Adjusting playback speed works best in combination with the SaneAR or MPC Audio Renderer. These two renderers have automatic pitch correction.” [Readme at tag 2.8.1] — and the newer release notes spell out the flip side: ”These audio renderers have automatic pitch correction, while DirectSound does not.”
The two settings you need
- View → Options (or press
O) → Playback → Output → Audio Renderer dropdown → select MPC Audio Renderer (or SaneAR Audio Renderer). If it currently says System Default / DirectSound, that is exactly the cause of your chipmunk effect.- Options → Playback → Speed step → set it to 10%. The rate buttons take a 2× step by default, which is why speed changes feel so coarse [Readme].
Then close and reopen the file (the DirectShow graph has to be rebuilt with the new renderer), and use
Ctrl+Up/Ctrl+Downto change rate,Ctrl+Rto reset.Since 2.8.0, the individual renderer settings live behind the small button next to the renderer dropdown on that same Output page, in case you want to inspect them.
Verifying it took effect
Open Play → Filters while a file is playing: the audio renderer entry should read MPC Audio Renderer, not DirectSound. If an external renderer (ReClock, an old sanear install, etc.) has been forced in the graph, that one wins and you’re back to varispeed.
Cases where it still won’t work
- Bitstreaming / passthrough (S/PDIF or HDMI with AC-3/DTS): a compressed bitstream can’t be time-stretched, so the renderer must receive PCM. Disable passthrough in the audio decoder if you use it.
- External audio renderers that don’t implement rate correction — same problem as DirectSound.
- Very high rates: DirectSound caps out around 4×, another symptom people hit before switching renderers.
If you’d rather not fight DirectShow at all, mpv does the right thing out of the box (
--audio-pitch-correctionis on by default and insertsscaletempo2), and VLC works as long as Enable time-stretching audio is checked.Minor note: 2.8.1 is from August 2026; the current build is 2.8.5, and 2.8.2 added things like showing playback rate in the status bar for audio files.
It’s right!
Not only it works, but the current stable version is 2.8.5.
But something is missing from its answer:
Between MPC Audio Renderer and SaneAR Audio Renderer, which one is better?
❸ Qwen4-Max:
Short version: for what you’re doing, they’re equivalent — both apply pitch correction. Take MPC Audio Renderer, and keep SaneAR in your back pocket as the fallback if a file sounds harsh.
Where each one comes from
- SaneAR is alexmarsev’s sanear renderer, absorbed into MPC-HC years ago. The upstream project’s last release was 0.3 in 2015.
- MPC Audio Renderer is the MPC-BE renderer, ported into MPC-HC in September 2021 (commit ”port of mpc(be) audio renderer (#1239)”) and still receiving work — the 2.7.2 and 2.7.3 changelogs from mid-2026 both list ”Updated MPC Audio Renderer”.
A doom9 regular’s summary matches that trajectory: ”there were troubles in the far long past with mpc audio renderer but for like many years now it pretty much supersedes sanear”. Meanwhile K-Lite’s author, asked exactly your question, answered: ”Both are good and are very similar in how they work and what functionality they offer” — with SaneAR being the pack’s default at the time.
The differences that actually exist
These come from clsid2 debugging a noisy sample against both renderers:
SaneAR MPC Audio Renderer WASAPI shared output format always 32-bit float passes the source format through (bit-exact oriented) Overflow handling has a limiter that reduces overall volume when samples exceed range no limiter — hot/clipping sources can audibly clip Maintenance upstream frozen since 2015 actively updated WASAPI exclusive + bitstreaming yes yes Two caveats worth knowing: exclusive mode is used implicitly whenever you bitstream, regardless of the checkbox; and MPCAR has had its own regressions, e.g. bitstreaming breaking when Windows 11 switches into HDR mode. They also share bugs — both choked on the same AVI with a zero-size ADPCM stream — which is unsurprising given the family resemblance.
I found no documented difference in the time-stretching itself, so don’t expect one to sound better than the other at 1.1×.
What I’d pick in your case
- Speech at 110–125%, stereo, shared mode → MPC Audio Renderer. Newer code, gets fixes, and the developer himself points people to it when DirectSound misbehaves [[28]].
- If a file sounds distorted/clipped (especially with volume boost or normalization enabled) → try SaneAR; its limiter is precisely what masks overflow in that scenario [[34]].
- Bitstreaming to an AVR → either; if one acts up on your setup, the other is the standard workaround.
- You installed via K-Lite and it’s already on SaneAR with no problems → the pack author’s advice stands: no need to change it.
One UI note
The renderer-specific options aren’t on the Output page itself: since 2.8.0 they sit behind the small button next to the renderer dropdown. You can also reach them during playback via Play → Filters → [your audio renderer] → Properties — that’s where you’d set the device, shared vs. exclusive, and bit-exact output.
For casual speed-watching, leave exclusive mode off: it locks the device against other applications and buys you nothing for pitch-corrected speech.

Leave a Reply