This post is mostly about how surprisingly hard it is to cut filler words out of speech cleanly. Apparently, stripping ums isn't a find and replace type thing, because Whisper's timestamps are off by up to a few hundred ms and cutting on them chops syllables or leaves stutters. So, I built a tool, erm, that starts from Whisper's guess, finds where each word actually starts and stops in the audio, and snaps the cuts to silence so there's no click, with ffmpeg doing the splicing.
What an awesome tool and idea. I’d be keen to see if it can integrate with video editing tools.
Ideally it would slice the video in the timeline without actually removing anything, so you can scrub through your video and try with and without each disfluency (thank you - awesome word) & decide case by case which to keep!
Really cool stuff and definitely going to try it; I’m also finding it wild that Google put effort into adding ums and erms into their text to speech model a while back. AI puts it in, AI helps take it out.
Not to promote something, but Wispr Flow does that for me automatically if I trigger a setting for it..
While it's a commercial product with a subscription, I spent a long time on the free tier not even hitting their limits until I started using it so extensively that I wanted to pay for it.
And I've used Whisper in the past, mostly for tinkering. I tried it for a couple of use cases but haven't touched the base project in a while. But I do regularly use Faster-Whisper-XXL, an open source project based on Whisper, for subtitle generation.
Though, for subtitle generation, I decided to support the project and mainly use the non-public build of Faster-Whisper-XXL Pro built for donators to the open source project.
The extra features smooth out the subtitle editing process very substantially. Toss in "--roformer_overlap 0.125 --roformer_vram 16 --best_of 15 --ff_vocal_extract mb-roformer --vad_method pyannote_v3" to the cli parameters (and sometimes --realign) and you have much less work to do in SubtitleEdit or Tero Subtitler afterwards to clean it up.
Surprisingly, it's the whisper model itself that does that. I find that it's also good with false starts, often correcting something like: "uhm, we could...we can go there" to just "we can go there", if spoken rapidly enough.
It’s a nice engineering approach, but I’m interested in the motivation. Um and ah is distracting in a transcript, where you can naturally pause to take in information; in speech however it can serve as a focusing point to indicate the next part is important. See https://medium.com/better-humans/dont-worry-about-saying-um-... for example. The weirdly obsessive zeal that orgs like Toastmasters have about eliminating them is weird.
Disfluencies aren’t necessarily bad even if the word starts with “dis”!
> in speech however it can serve as a focusing point to indicate the next part is important
it's... exact opposite?
the main (attempted) use for ummms is to keep continuation of speech despite the pause. And the main complaint is exactly that it ruins the focus and doesn't give respite
The most popular academic theory (IIRC) is that "um" and "uh" are conversational placeholders that say, "don't talk, I'm not finished speaking yet". Which obviously serves no purpose in a monologue.
To me they just indicate lack of confidence on the part of the speaker.
The younger generation seems to love listening at 1.2x or faster. I think it’s a preference for a fast information dopamine hit. I may argue it’s even a shallow approach that prefers against pausing and time for careful reflection. Meanwhile, book reading is at an all time low seemingly because no one has a preference or patience for careful study and reflection.
You are almost certainly listening to a lot of audio with um's and ah's removed from it, without even realizing it. Via various methods including automated, manual, direction (re-takes), and talent-led (speaker consciously suppressing).
If you knew how much, you might not feel the same way.
I would love to see support for videos and removal of custom filler words (I say 'basically' and 'like' a lot and have so far failed to improve myself on this).
This approach seems kind of backwards to me. Why try to detect everything except the thing you're trying to remove instead of either sampling a few uhs and ums and treating them as noise to be silenced (with a sharp crossfade to the noise floor that doesn't interrupt speech flow) or finetuning a model to detect them specifically for full automation?
> instead of either sampling a few uhs and ums and treating them as noise to be silenced
If you're not paying ttention, ctting out specific sounds can easily cause more trouble. I for one would be quite pset if I couldn't hear the pire's reasoning for calling a foul.
I think it is harder to remove those from your own speech. I have been doing that for few months now and I still get back at it when I am in hurry or stressed.
In my experience native English speakers are particularly bad, generally when speaking a second language people are less likely to add random filler words.
Also the type of filler word for some reason is often different between UK and US: British people tend to be "umm"-ers and Americans are more likely to add "you know" (although "umm" is also common).
Once you notice it it's impossible to ignore and many, many native English speakers are actually terrible at speaking and add filler words to the point where it's very distracting
This is great, I've tried out automated podcast editing tools before and they cut too aggressively in my experience. What are you thinking about doing next with this now that you've gotten the alignment snapping working cleanly for 'um' and 'ah', are you thinking of expanding the tool?
I think the “What it won’t touch” section shows why the entire concept is unsound. Here it is with a different first sentence, and (other than the third sentence no longer matching erm’s reality) it’s perfectly coherent:
> It leaves um, uh, er and elongated versions (ummmm, uhhhhh) alone. Those sound like fillers but they’re doing real work in the sentence, and cutting them automatically would change what someone said. The rule erm follows: only remove things that are sound, not language.
> It also doesn’t touch repeated words, false starts, or long thinking pauses. Those aren’t noise on top of the speech; they are the speech, just messier than the speaker would like. Cleaning them up is an editorial decision about which take to keep, and erm doesn’t have an opinion about that.
Think about it. Cleaning these things-that-can-be-just-sounds-but-can-also-very-much-be-load-bearing up is an editorial decision. At the very least, you need to judge based on the surrounding content whether the removal of an um would change the meaning at all; and I don’t think text alone is adequate for that.
I wonder if with enough input data and transcription you could “fingerprint” where a speaker personality has habits of interjecting “ums” leading to more hardy analysis. Novel approach, but gets me thinking
When I was doing podcasts regularly, it made me acutely aware of various people's speech mannerisms. (Somewhat similarly, recording a lot of videos during COVID made me very aware of a variety of my own mannerisms--especially overactive hand motions.)
> Two small fixes, in order. First, each cut endpoint is allowed to slide a tiny bit (up to 60ms) to land in the quietest spot nearby. If there’s a momentary lull in the audio just before or after the original cut point, slide there. The slide is bounded so it can’t cross into a neighboring word, otherwise you’d chew off real speech. Second, from that quiet spot, the endpoint snaps to the nearest moment when the waveform is exactly crossing zero.
No, you run an entire second pass LLM over the output of Whisper. "no uhhh three no four." should just output four the numeral not even f.o.u.r.
Hi, my name is fragmede. Judging by the date on my computer it's been four months since it's since I've t touched the transcription directory on computer and tried to improve on the state of wisprflow. Mines pretty good but it just doesn't... ah you can't drag me back in.
46 comments
[ 3.9 ms ] story [ 69.4 ms ] threadhttps://github.com/dougcalobrisi/erm
Ideally it would slice the video in the timeline without actually removing anything, so you can scrub through your video and try with and without each disfluency (thank you - awesome word) & decide case by case which to keep!
A trivial example is "umm... well... (sigh) okay" versus just "okay". Not okay!
While it's a commercial product with a subscription, I spent a long time on the free tier not even hitting their limits until I started using it so extensively that I wanted to pay for it.
And I've used Whisper in the past, mostly for tinkering. I tried it for a couple of use cases but haven't touched the base project in a while. But I do regularly use Faster-Whisper-XXL, an open source project based on Whisper, for subtitle generation.
Though, for subtitle generation, I decided to support the project and mainly use the non-public build of Faster-Whisper-XXL Pro built for donators to the open source project.
The extra features smooth out the subtitle editing process very substantially. Toss in "--roformer_overlap 0.125 --roformer_vram 16 --best_of 15 --ff_vocal_extract mb-roformer --vad_method pyannote_v3" to the cli parameters (and sometimes --realign) and you have much less work to do in SubtitleEdit or Tero Subtitler afterwards to clean it up.
Disfluencies aren’t necessarily bad even if the word starts with “dis”!
it's... exact opposite?
the main (attempted) use for ummms is to keep continuation of speech despite the pause. And the main complaint is exactly that it ruins the focus and doesn't give respite
To me they just indicate lack of confidence on the part of the speaker.
If you knew how much, you might not feel the same way.
I might add the custom filler word functionality and/or perhaps just make the filler word list configurable.
If you're not paying ttention, ctting out specific sounds can easily cause more trouble. I for one would be quite pset if I couldn't hear the pire's reasoning for calling a foul.
Also the type of filler word for some reason is often different between UK and US: British people tend to be "umm"-ers and Americans are more likely to add "you know" (although "umm" is also common).
Once you notice it it's impossible to ignore and many, many native English speakers are actually terrible at speaking and add filler words to the point where it's very distracting
> It leaves um, uh, er and elongated versions (ummmm, uhhhhh) alone. Those sound like fillers but they’re doing real work in the sentence, and cutting them automatically would change what someone said. The rule erm follows: only remove things that are sound, not language.
> It also doesn’t touch repeated words, false starts, or long thinking pauses. Those aren’t noise on top of the speech; they are the speech, just messier than the speaker would like. Cleaning them up is an editorial decision about which take to keep, and erm doesn’t have an opinion about that.
Think about it. Cleaning these things-that-can-be-just-sounds-but-can-also-very-much-be-load-bearing up is an editorial decision. At the very least, you need to judge based on the surrounding content whether the removal of an um would change the meaning at all; and I don’t think text alone is adequate for that.
- Ums Considered Harmful: https://hamanlp.org/research/ums/
- Related paper: https://hamanlp.org/SIGBOVIK_2026.pdf
for something dealing with audio you do need to play the audio really
Oh, Claudish striking again.
No, you run an entire second pass LLM over the output of Whisper. "no uhhh three no four." should just output four the numeral not even f.o.u.r.
Hi, my name is fragmede. Judging by the date on my computer it's been four months since it's since I've t touched the transcription directory on computer and tried to improve on the state of wisprflow. Mines pretty good but it just doesn't... ah you can't drag me back in.