Repository navigation
Replies: 25 comments 24 replies
|
Great initiative @ryanheise Here is one to start - I have been looking at creating short synthetic audio clips that induce hallucination and have no private or copyrighted content. With |
|
I frequently get these repeats like this, requiring the need to re-do them. |
|
The purpose of this thread is merely to share examples of hallucinations so that people working on solutions to that problem can test their solutions on your examples. Do you have any examples of hallucinatory audio data to share? |
|
Oh sorry, my apologies. |
|
@ryanheise , do I understand right that we can use those audio segments that produce hallucination lines to re-train (finetune) Whisper model? Is that what you meant? [duplicate warning]: I had just posted a similar note in another thread before I saw this conversation. |
|
There are a variety of ways people are already trying to solve the hallucination problem, but in order to measure how effective those solutions are, or to compare one solution with another, we would need data to test it on. So the purpose of this thread is just to collect data. Do you have any hallucinatory audio samples you could share? |
|
I like this one, it has given different hallucinations each run. And I had never heard of a Cyclophone before but apparently it's a real thing. :)
Run 1 Run 2 Run 3 large-v3 consistently seems to produce the same (less interesting) hallucination on this audio |
|
@ryanheise This file also triggers hallucination in other languages, and the transcript also seems to be different each time. So this gives the possibility of a single audio file giving a much bigger set of hallucination data to evaluate. Can I assist here by generating multiple whisper outputs from this one file (ie. multiple runs across multiple languages)? Is .json output helpful here or do you need to run something custom? |
|
That's a good question since my top post was a little brief (I've edited it) - the main things are 1) the audio file and 2) which timestamps specifically to look out for hallucination, because if someone is working on a solution, they will also be running the inference process many times themselves and will be able to see the full output, perhaps with their own additional debug output. Those would be the main things, but obviously anything else you might share in addition to that can possibly be interesting and valuable too. For example, when I ran whisper on your file, it seems I got very different hallucinations to yours. I've only run it twice, but in both cases, I got "Thanks for watching! Applause" and "Thank you for watching! Applause". I'm not sure if that's maybe because you were running some other options in addition to |
|
Ok, well the .json will record the timestamps, the no_speech probabilities, etc so hopefully that works. But more importantly now you have me wondering now about the reproducibility of different hallucinations each run since you got different results. For initial testing I ran with just I ran this again just now and got 3 different hallucinations in 3 runs. I also confirmed that I get different hallucinations on Colab, using large-v2 on CUDA and lang='es' as shown below. My thought is to use Colab to generate a large number of .json hallucinations, if you think that will be helpful. result = model.transcribe(vfile, language=lang, fp16=False, beam_size=5, task='transcribe',
best_of=5, word_timestamps=True, verbose=True) |
|
...and after retesting on CUDA with language en, I got this Note that this is for the second posted file, 5-2. |
I was thinking more along the lines of where, if someone shares a file that tends to induce hallucinations, but the file is 1 hour long, and the hallucinations tend to occur at the beginning, and at the 18 minute mark, then sharing those two locations will be very helpful. Hopefully that sets the bar low on contributing, since it will be better to collect a wide variety of different audio files with a variety of different trigger conditions for hallucinations. |
|
Why does the Whisper return random text (hallucinations) for an audio that does not contain any speech? I tested it with random traffic and one minute of silence. The output contains few words. But as per my thinking, it should not return anything in case there is no speech in the audio. How can I handle this ? Output: Output: |
|
@qaixerabbas thanks for sharing the samples.
It is likely because the training data contained transcripts during silence. For example, if you think about subtitles for movies, sometimes you might see a Copyright notice in the subtitles, or you might see credits such as "Subtitles written by xyz", even though there is no actual speech happening in the audio at the timestamp where these subtitles appear. So imagine the neural net learning from this data. It will basically learn (for example) that when there is silence at the end of the video, it should "make up" some text to go along with the silence, the sort of text that it learns typically goes at the end of a video. The situation would be better with carefully cleaned up training data, although there is a tradeoff between how much effort it takes to clean up data, and how massive a data set you can build (more data = better). There are many solutions and workarounds people are working on to prevent hallucinations, and most of them involve just cutting out the silent parts of the audio so that Whisper is never tempted by the silence (you can search the discussions for "hallucinations" to find the various solutions). |
|
@ryanheise thank you for the detailed explanation. I will look at this. |
|
So I guess no one here has heard about the --suppress_tokens "" trick? I've posted about it before, maybe it was simply lost in traffic. UPDATE: This method does NOT work with Turbo (and by extension, it probably won't work with large-v3 either). |
The purpose of this discussion is to work together to build a dataset of audio examples that induce hallucinations so that people can test their solutions on it. (For example, you could test your trick/solution on these 4 examples to see how it compares against other solutions.) |
|
Right, I don't mean to derail this too much. These samples are very useful, thank you anyone who donated :D EDIT: Here's what I got with the synthetic samples (this is with Whisper version 20230314):
|
|
@misutoneko Thank you for the comments and for trying the synthetic samples. With the 5-2 sample, I had been getting wildly different results each run. Using If I had to guess, what is happening here is that the default list of suppression tokens causes the identification of background sounds (typing, music, applause, etc) to be ignored - and decoding is allowed to continue until it hallucinates speech. |
|
Yes, it seems like it's a combination of bad training data and the way the default suppression works. The prediction is kind of "forced to take a wrong turn". Note that hallucinations happen in whisper.cpp too, but in a different way because they don't use the same code. Here's my original thread from July: |
|
Hi @misutoneko , do the tokens need to be grouped/sequenced together when passed to --suppress_tokens? For example below is a hallucination. Should the command be --suppress_tokens "50364, 7, 6847, 96, 7016, 32045, 8, 50464" ? Does this suppress any occurences of these tokens in results, or only if they are sequenced as given? |
|
AFAICR there was a bug in command line handling of this parameter (only one token can be given), so it's either --suppress_tokens "" (as in, empty string) which means that nothing is suppressed. Or alternatively --suppress_tokens 50364 or --suppress_tokens 50363 depending on whether you have english-only or multilingual model. That will suppress the BEG token only, and that also worked for me with whisper-timestamped (I haven't tried it with regular/vanilla Whisper). It probably induces the "no timestamp" mode, but that's not necessarily a problem (if there's no actual speech in the clip, you probably don't care about the timing). If you really want to give multiple tokens, I think that can be done via Python, but I'm not so sure that's a good idea. |
|
Hi, here is a 3min audio with a combination of music and kitchen noise. The hallucination liness by large-v2: And, here is the hallucination by large-v3: And, hallucination by large-v1: |
|
With |
|
FWIW, my shop has processed thousands of hours of English files with a pretty wide range of speakers and lots of silence/music hallucinations. (we stick to large-v2). All our outputs are VTTs which we post-process by manipulating the output with webvtt CaptionLists. I do dozens of other corrections for casing, time, named entities, and line length that are all particular to my org's use cases (currently in a private repo, sorry I can't share the whole thing ATM), but here's a handful of the things that are most common in what we see, YMMV obviously: We also filter to a limited subset of unicode characters to drop obvious non-latin strings. Rationale: Why would Asiatic characters ever legitimately appear in an English text? Additionally:
None of these are prefect but they cut down a lot of the extra noise to a more reasonable looking output in many circumstances. |
|
Hi @whicks1 , have you seen #928 (comment) ? You might want to share your work with everyone there since that discussion is building a set of text outputs rather than the audio files themselves. |
|
Now, I want to download the top XF SUPER SUS TOWER
|
|
First time I've seen this hallucination "KING OF THE R" during outro music of segment. Unless of course the outro music is called "KING OF THE R" and it is being "labelled" What do I mean by "labelled"? When a segment is announced, and then it repeats the name of the segment: "The Wooden Prince by Bela Bartók is played now by Christian Maciolaru." (source https://www.bbc.co.uk/programmes/m001whrr ) Note that after that it completely goes off the rails. Where are the hallucinations of form "StSq2 2.60" and "3F 5.83 x (-1.00), FCSp4 3.20" coming from?? |
|
Another fun hallucination is "counting" (source: https://www.bbc.co.uk/programmes/m001qklp ) Source: https://www.bbc.co.uk/programmes/m001rws1 also "En-za-mex" == "In the mix" ?? |
|
FYI, although I'm not locating it right now, I've seen a "counting hallucination" go to near 100. In correct sequence, etc. Which is quite an emergent "intelligence" of the model. How does it know to count, and why does it count during hallucinations? |
|
https://archive.org/details/movie-1992-musketeers.twenty.years.after Whisper surely loves you. Specially when playing |
|
Here's one i just got - searched "openai CastingWords" on google and found this thread. "Transcription by CastingWords" Wtf? |
|
As explained in the initial post, can you please include:
|
|
Hi, I encounter a hallucination problem when using Whisper-large-v3. Please see here. |
|
Hello, I am working on Whisper-large-v3 to transcribe Chinese audio and I meet the same hallucination problem as #2071. |
|
Can you share your mp3 file here? |
|
Here is the mp3 and complete result, thanks! { |
|
which VAD method did you use? |
|
So I finally figured out where this class of hallucination comes from -- figure skating: https://www.isuresults.com/results/season2425/jgpcze2024/FSKXPAIRS---JUNIOR----QUAL000100--_JudgesDetailsperSkater.pdf
(source https://www.bbc.co.uk/programmes/m0022ctz ) Suggestion: don't use train your language models on Ice Skating videos, or classical music will trigger transcription of figure skating notations. :-) |
|
I met the same issue of last post mentioned when transcripting Chineese audio file, it happened at the end of the sound file. Both the "好" and the "字幕志愿者 杨茜茜" are hallucinations. |
|
May I ask which model did you use? |
|
Is this a kind of hallucination? Whisper generated the wrong subtitle for the video: |
Background noise, |
|
Same with turbo-v3. Here's a clip of one of the hallucinated locations (cut from 11s..to 17s)
|
|
Source: https://www.bbc.co.uk/programmes/m002q42l
|
|
Spurious repetitions are still too frequent. I would like to share a workflow in which I tackled them by creating an optional parameter named |
|
maybe this is useful for someone here still: |
|
Two things we learned running Whisper in production on formal French dictation (legal reports, where a wrong word is a liability), in case they help others here. Most of our phantom phrases were our own fault. Our pipeline cut the audio at every photo timestamp so that each photo could be anchored to a segment. Between two bursts of photos that produced 1–5 s slices with no context, and those are where "Merci d'avoir regardé cette vidéo", "Sous-titrage Société Radio-Canada" and "This is it." showed up: forced short segments hallucinated roughly nineteen times more often than the rest, and dictations without photos (hence without forced cuts) had about twenty times fewer phantom phrases per hour of audio. The engine's own options ( What fixed it: keep the cuts, but give the model ~10 s of audio on each side of any segment under 10 s, then keep only the words whose midpoint falls inside the original segment (word timestamps). Boundaries never move. Phantom phrases nearly vanished, quality went slightly up, coverage unchanged, about a sixth more compute. One trap: turn The phrases follow the decoded language, not the expected one. A French dictation that drifts returns English credits. We keep per-language lists (en/fr/de/es/it/pt) and check all of them. The detectors, the context-window logic and the write-ups (including the variants that looked better and were refused, e.g. "no cuts at all" silently dropped almost a tenth of the audio) are open source, MIT: https://github.com/Guillain-RDCDE/Trusted-Transcription — there is an in-browser demo where you can paste a transcript and see what gets flagged: https://guillain-rdcde.github.io/Trusted-Transcription/ Happy to add phantom phrases you have observed in other languages to the lists. |
|
This isn't necessary a hallucination, but when whisper have overlapping sound effects/voices, whisper have serious problems. I would expect something like "[multiple overlapping voices]", but instead got this (after "We're going to get married.") and the rest of lines are out of synch. A public source to the problematic location in movie |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Please share below any links to audio/video files that you have found to induce hallucinations in Whisper.
You may include:
The people working on various solutions to the hallucination problem can use these examples to help evaluate and improve those solutions.
All reactions