word-level timestamps in `transcribe()` by jongwook · Pull Request #869 · openai/whisper

jongwook · 2023-01-20T20:31:43Z

No description provided.

ryanheise · 2023-01-21T00:32:58Z

This DTW dependency introduces a licence incompatibility, but an alternative was suggested earlier in the discussions from memory.

Edit: Alternative library recommended in #813 (reply in thread)

ryanheise · 2023-01-21T00:42:04Z

whisper/transcribe.py

+        segment["words"] = []
+
+    for i, (word, start, end) in enumerate(zip(words, start_times, end_times)):
+        if word.startswith("<|") or word.strip() in ".,!?、。":


I think you're missing ？ and others. See:

https://github.com/jianfch/stable-ts/blob/5fdaff51caec148c80b8e70311e8c62d07b07146/stable_whisper/stabilization.py#L440-L443

Would actually be convenient to actually re-insert the punctuation tokens so that concatenating all the words is the same as concatenating all the tokens. That would just make processing easier on the consumer end. For reference, Amazon Transcribe includes timestamped punctuation tokens in the results.

I notice there is now a TODO comment leaning to the first approach:

if word.startswith("<|") or word.strip() in ".,!?、。": # TODO: expand

I'm not sure if you've already committed to that approach, but I would vote for not removing the punctuation, so that whether a consumer wants to traverse the entire content by token, by word, or by segment, they can and do it in either of these 3 ways and in all the content is there (the concatenation of each result is identical). Otherwise if I consume the results by word, I would need to simultaneously look up one of the other two results to cross reference the, and look for the bits that were omitted from the sequence. Here is how Amazon Transcribe does it, for example.

On the other hand, if that is not persuasive, you might consider instead making it an option whether or not to strip out the punctuation.

(I note also that if you just left the punctuation in, the consumer would still have the ability to filter them out.)

Thanks for the suggestions; I've updated so that the punctuation characters are attached to the adjacent words, while keeping the word's timestamps.

Thanks, that looks good. I think if the prepend_punctuations and append_punctuations parameters were propagated in transcribe() and cli() that would be quite helpful, since then I could set them to empty strings to emulate the Amazon Transcribe behaviour.

KaiserChr · 2023-01-21T08:30:22Z

Hi!
I tried out this branch with kwargs['word_level_timestamps'] = True but the model performed very slowly. In addition (or rather because of) it started to hallucinate like mad.
Im using chunks of short (couple of seconds) audio data in german produced by a VAD for live transcription.

Maybe its a problem on my side, maybe anyone can try to reproduce?

jongwook · 2023-01-21T08:35:38Z

Thanks for the comments, all -- this is work in progress and not quite ready for merging. I'm trying to address both hallucination and performance concerns.

glangford · 2023-01-24T01:37:16Z

Yet another DTW implementation, fyi. Can't vouch for it other than to say that it is Apache licensed, recently updated, has both pure Python and C implementations.

https://github.com/wannesm/dtaidistance

Jeronymous

That feature of getting word-level timestamps is useful for plenty of applications. Excellent work!

EDIT
Question: Would it be possible to hook cross-attention weights on the fly, while the model is transcribing, and avoid doing inference with the decoder twice?
It would be more efficient, and would maybe better take into account factors like the context (conditioning on previous text...).

Jeronymous · 2023-01-24T22:44:07Z

whisper/timing.py

+            tokenizer.timestamp_begin + mel.shape[-1] // 2,
+            tokenizer.eot,


I don't thing these last two tokens are needed to estimate word timestamps.

This was a trick to nudge the DTW path to go along these tokens so that the last few words have more accurate timestamp. It's still not perfect, but I settled with using <|no_timestamps|> token and no timestamp tokens in the recent commit.

I understand that it's important to have attention weights used to predict the timestamp token for the end of speech segment, but these attention weights are the one you get when the input is the last predicted (sub)word token. I think it's enough. When the input token is the final timestamp the decoder is already focusing on predicting the next thing.
I wonder if things are not shifted by one, because it was a problem I saw with your notebook (the timestamps were assigned to the token before the one it should be).

We are anecdotally seeing that too, in our tests. The timestamps lag a word for example Got no empirical proof ( very anecdotal ).

I don't see the same observations.

I did notice in earlier commits that the next token after a comma may lag as if the comma was taking up too much time. That seems to have become more accurate in later commits.

hmm, we could not come with any empirical evidence either. May it was the previous version.

Jeronymous · 2023-01-24T22:44:30Z

whisper/timing.py

+    for hook in hooks:
+        hook.remove()
+
+    weights = torch.cat(QKs[-6:])  # layers * heads * tokens * frames


Why only considering (at most) the last 6 layers?

This was because the attention weights in the later layers were more indicative of the time alignment. I've updated this part, and now it uses a mask to select which layers and heads to find the alignment.

Ok interesting. I have to check this masking trick.
I don't understand why later layers are more indicative. Is it an intuition that I am missing, or some empirical results you got from experiments?

Jeronymous · 2023-01-24T22:47:14Z

whisper/timing.py

+            continue
+
+        segment = segments[token_sources[word_boundaries[i]]]
+        segment["words"].append(dict(word=word, start=round(start, 2), end=round(end, 2)))


Would it be possible to add a confidence score based on the average log proba for each words?
This can be a useful feature, available with very little additional computations.

Great point! Added in 5fa4356

Awesome 👍

Jeronymous · 2023-01-24T22:54:50Z

whisper/timing.py

+            c1 = cost[i - 1, j]
+            c2 = cost[i, j - 1]


Maybe one of these two is not needed, as it doesn't really make sense to attribute a same timestamp to several tokens.
Well... not fully sure. Maybe it's useful when a lot of text has to be aligned with a small portion of audio (which can happen when Whisper "inner language model" is stuck).

Yeah it totally makes sense to force a token to have at least one timestamp, which is only about 20 milliseconds. I left this as-is, to handle some failure cases like repetition looping as you mentioned; in the post-processing zero-length segments are removed, and it was usually the case for the generation got stuck on repetition looping.

We have seen words repeating with same end and start timestamps. Segments were fine. I think we have to move the segment code where you handle duplication ( repetition looping ) to words too.

taylorchu · 2023-01-27T05:22:50Z

whisper/triton_ops.py

+
+
+@triton.jit
+def dtw_kernel(cost, trace, x, x_stride, cost_stride, trace_stride, N, M, BLOCK_SIZE: tl.constexpr):


triton jit requires gcc/clang, python-dev, and cuda-dev at runtime. Please consider some lighter-weight alternatives.

Thanks for the heads up; I made it fall back to pytorch and numba implementations if triton fails with RuntimeError or subprocess.CalledProcessError. I haven't tested this on a non-dev environment, so please feel free to ping me if the fallback does not work for any reason.

+1. Thank you.

saunair · 2023-02-23T17:42:17Z

whisper/timing.py

+
+    # heads * tokens * frames
+    weights = torch.stack([QKs[l][h] for l, h in model.alignment_heads.indices().T])
+    weights = weights[:, :, : num_frames // 2]


Question: Why get weights only for the first num_frames/2 frames?

saunair · 2023-02-23T17:45:09Z

tests/test_timing.py

+
+
+@pytest.mark.parametrize("N, M", sizes)
+def test_dtw(N: int, M: int):


Was your reason to not use the dtw library licensing concerns or just speedup?

dtw-python is GPL, as mentioned here -
#869 (comment)

saunair · 2023-02-23T17:46:59Z

tests/test_timing.py

+
+
+@pytest.mark.parametrize("shape", shapes)
+def test_median_filter(shape):


Question: Is there a licensing issue using scipy.median_filter or is this cuda implementation just faster?

ryanheise · 2023-02-25T07:58:28Z

I found an interesting edge case with the small model where enabling the word-level timestamps option causes it to repeat the prompt at the end of the audio while also failing to infer the last word.

$ ffmpeg -t 29 -i https://audio2.redcircle.com/episodes/6b196013-8672-43d9-be52-4332b3207d93/stream.mp3 test.mp3

$ whisper --model small test.mp3
.../whisper/transcribe.py:98: UserWarning: FP16 is not supported on CPU; using FP32 instead
  warnings.warn("FP16 is not supported on CPU; using FP32 instead")
Detecting language using up to the first 30 seconds. Use `--language` to specify the language
Detected language: English
[00:00.000 --> 00:15.920]  Military veteran Eric Weinstein began 69 Whiskey as a college radio show on 107.7 The
[00:15.920 --> 00:21.720]  Bronx, located on the campus of Ryder University in Lawrenceville, New Jersey.
[00:21.720 --> 00:27.560]  A show once restrained by rules and boundaries now comes straight to you raw, uncensored and
[00:27.560 --> 00:28.960]  unapologetic.

$ whisper --model small --output_format json --word_timestamps True test.mp3
.../whisper/transcribe.py:98: UserWarning: FP16 is not supported on CPU; using FP32 instead
  warnings.warn("FP16 is not supported on CPU; using FP32 instead")
Detecting language using up to the first 30 seconds. Use `--language` to specify the language
Detected language: English
[00:08.040 --> 00:15.940]  Military veteran Eric Weinstein began 69 Whiskey as a college radio show on 107.7 The
[00:15.940 --> 00:21.320]  Bronx, located on the campus of Ryder University in Lawrenceville, New Jersey.
[00:21.720 --> 00:28.980]  A show once restrained by rules and boundaries now comes straight to you raw, uncensored and
[00:28.960 --> 00:28.960]  Military veteran Eric Weinstein began 69 Whiskey as a college radio show on 107.7 The
[00:28.960 --> 00:28.960]  Bronx, located on the campus of Ryder University in Lawrenceville, New Jersey.
[00:28.960 --> 00:28.960]  A show once restrained by rules and boundaries now comes straight to you raw, uncensored and
[00:28.960 --> 00:28.960]  Military veteran Eric Weinstein began 69 Whiskey as a college radio show on 107.7 The
[00:28.960 --> 00:28.960]  Bronx, located on the campus of Ryder University in Lawrenceville, New Jersey.

IgnacioSan22 · 2023-02-27T09:35:59Z

Hi @jongwook ,
Since you first release the notebook to obtain word-level timestamps I've been working on this to add to whisper process. And I've tried to test other alingment methods than DTW. Have you tried something else and found out that it works better?

Also, I've been struggling a lot with alucinations, specially for spanish content. I've create a cleaner function at segmet level, is there any smarter way?

ioskevinshah · 2023-03-06T12:28:53Z

is there any chance to have word level timestamps in Whisper API?

jongwook · 2023-03-06T22:00:36Z

Hi @IgnacioSan22, the custom DTW implementation in this PR was for the license issue as noted by others and also for the speed. An alternative is to use the timestamp predictions from the model, but we found that it's less reliable than using the attention patterns like in this PR. If you have solutions using any other algorithms for alignment, please let me know!

The community had some success handling hallucinations by preprocessing the inputs with VAD, like:

Hi @ioskevinshah, this feature is still experimental but we do plan to add it to the API as an option, once we're sure that it's reliable enough.

JeffreyWardman · 2023-03-08T04:23:40Z

@jongwook is there a way to access it via a beta flag for instance? How can we know when something is/isn't added to the API?

jongwook · 2023-03-08T04:28:44Z

For the API, the speech-to-text guide and the audio API reference provide the full documentation of the available features. These documents will be updated accordingly as we roll out new features.

IgnacioSan22 · 2023-03-08T10:58:51Z

Hi @IgnacioSan22, the custom DTW implementation in this PR was for the license issue as noted by others and also for the speed. An alternative is to use the timestamp predictions from the model, but we found that it's less reliable than using the attention patterns like in this PR. If you have solutions using any other algorithms for alignment, please let me know!

The community had some success handling hallucinations by preprocessing the inputs with VAD, like:

A possible solution to Whisper hallucination #679

Whisper WebUI with a VAD for more accurate non-English transcripts (Japanese) #397

https://github.com/m-bain/whisperX

Hi @ioskevinshah, this feature is still experimental but we do plan to add it to the API as an option, once we're sure that it's reliable enough.

Hi @jongwook, I've tried the hungarian algorithm and in some cases the results are better, however due to the lack of resources I'm not capable to perform a proper study to find the best alingment algorithm. For hallucinations I've developed a postprocess functions that cleans the segments. It improves quite a lot, but I'll check those references.

Thanks

glinft · 2023-03-15T06:09:54Z

For the API, the speech-to-text guide and the audio API reference provide the full documentation of the available features. These documents will be updated accordingly as we roll out new features.

One more question: when will this new feature be rolled out?

ioskevinshah · 2023-03-23T13:18:23Z

Hi @IgnacioSan22, the custom DTW implementation in this PR was for the license issue as noted by others and also for the speed. An alternative is to use the timestamp predictions from the model, but we found that it's less reliable than using the attention patterns like in this PR. If you have solutions using any other algorithms for alignment, please let me know!

The community had some success handling hallucinations by preprocessing the inputs with VAD, like:

A possible solution to Whisper hallucination #679

Whisper WebUI with a VAD for more accurate non-English transcripts (Japanese) #397

https://github.com/m-bain/whisperX

Hi @ioskevinshah, this feature is still experimental but we do plan to add it to the API as an option, once we're sure that it's reliable enough.

any workaround or logic after the API response?

* word-level timestamps in `transcribe()` * moving to `timing.py` * numba implementation for dtw, replacing dtw-python * triton implementation for dtw * add test for dtw implementations * triton implementation of median_filter * a simple word-level timestamps test * add scipy as dev dependency * installs an older version of Triton if CUDA < 11.4 * fix broken merge * loosen nvcc version match regex * find_alignment() function * miscellaneous improvements * skip median filtering when the input is too small * Expose punctuation options in cli and transcribe() (openai#973) * fix merge error * fix merge error 2 * annotating that word_timestamps is experimental --------- Co-authored-by: ryanheise <ryan@ryanheise.com>

samuelbradshaw · 2023-06-15T02:46:47Z

This is awesome! Is there a way to pass in pre-transcribed text that whisper can use for more accurate alignment?

* word-level timestamps in `transcribe()` * moving to `timing.py` * numba implementation for dtw, replacing dtw-python * triton implementation for dtw * add test for dtw implementations * triton implementation of median_filter * a simple word-level timestamps test * add scipy as dev dependency * installs an older version of Triton if CUDA < 11.4 * fix broken merge * loosen nvcc version match regex * find_alignment() function * miscellaneous improvements * skip median filtering when the input is too small * Expose punctuation options in cli and transcribe() (openai#973) * fix merge error * fix merge error 2 * annotating that word_timestamps is experimental --------- Co-authored-by: ryanheise <ryan@ryanheise.com>

word-level timestamps in transcribe()

8f9357f

jongwook force-pushed the word-level-timestamps branch from 6a75e25 to 8f9357f Compare January 20, 2023 20:35

ryanheise reviewed Jan 21, 2023

View reviewed changes

jongwook and others added 3 commits January 21, 2023 01:15

moving to timing.py

46ea501

Merge branch 'main' into word-level-timestamps

cfd2b81

numba implementation for dtw, replacing dtw-python

742d2f4

jongwook force-pushed the word-level-timestamps branch from e5e45ec to 742d2f4 Compare January 22, 2023 05:50

jongwook and others added 3 commits January 22, 2023 00:54

Merge branch 'main' into word-level-timestamps

fb12414

triton implementation for dtw

80331c0

add test for dtw implementations

1d2ed66

jongwook and others added 5 commits January 24, 2023 09:30

triton implementation of median_filter

b61e8f4

a simple word-level timestamps test

54f2901

add scipy as dev dependency

8ce6277

Merge branch 'main' into word-level-timestamps

812f446

installs an older version of Triton if CUDA < 11.4

cd5191f

jongwook force-pushed the word-level-timestamps branch from e325fa8 to cd5191f Compare January 24, 2023 20:07

jongwook and others added 3 commits January 24, 2023 12:08

Merge branch 'main' into word-level-timestamps

f64d8bc

Merge branch 'main' into word-level-timestamps

89133bd

fix broken merge

d4f9399

Jeronymous reviewed Jan 24, 2023

View reviewed changes

jongwook and others added 3 commits January 24, 2023 14:59

Merge branch 'main' into word-level-timestamps

040aa04

loosen nvcc version match regex

8e2756b

find_alignment() function

6c431c4

taylorchu reviewed Jan 27, 2023

View reviewed changes

jongwook and others added 2 commits February 2, 2023 04:04

Merge branch 'main' into word-level-timestamps

ff6cbfd

miscellaneous improvements

5fa4356

jongwook mentioned this pull request Feb 16, 2023

Expose punctuation options in cli and transcribe() #973

Merged

Expose punctuation options in cli and transcribe() (#973)

8eb29c3

Pikauba mentioned this pull request Feb 22, 2023

Add per token confidence to each segment. #991

Draft

guillaumekln mentioned this pull request Feb 23, 2023

Shorter segments? SYSTRAN/faster-whisper#15

Closed

saunair reviewed Feb 23, 2023

View reviewed changes

jongwook and others added 4 commits March 6, 2023 16:02

Merge branch 'main' into word-level-timestamps

6ed4c11

fix merge error

31cd418

fix merge error 2

145f325

annotating that word_timestamps is experimental

2b079c4

jongwook merged commit 500d0fe into main Mar 6, 2023

jongwook deleted the word-level-timestamps branch March 6, 2023 22:01

ryanheise mentioned this pull request Mar 8, 2023

attempt to fix the repetition/hallucination issue identified in #1046 #1052

Merged

bofenghuang mentioned this pull request Dec 28, 2023

[Whisper] Add word timestamps and confidence scores ml-explore/mlx-examples#201

Merged

ZachNagengast mentioned this pull request Feb 16, 2024

Word level timestamps argmaxinc/argmax-oss-swift#2

Closed

		tokenizer.timestamp_begin + mel.shape[-1] // 2,
		tokenizer.eot,



		@triton.jit
		def dtw_kernel(cost, trace, x, x_stride, cost_stride, trace_stride, N, M, BLOCK_SIZE: tl.constexpr):



		@pytest.mark.parametrize("N, M", sizes)
		def test_dtw(N: int, M: int):



		@pytest.mark.parametrize("shape", shapes)
		def test_median_filter(shape):

Conversation

jongwook commented Jan 20, 2023

Uh oh!

ryanheise commented Jan 21, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ryanheise Jan 21, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

KaiserChr commented Jan 21, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jongwook commented Jan 21, 2023

Uh oh!

glangford commented Jan 24, 2023

Uh oh!

Jeronymous left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

saunair Feb 23, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

saunair Feb 23, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ryanheise commented Feb 25, 2023

ryanheise commented Jan 21, 2023 •

edited

Loading

ryanheise Jan 21, 2023 •

edited

Loading

KaiserChr commented Jan 21, 2023 •

edited

Loading

Jeronymous left a comment •

edited

Loading

saunair Feb 23, 2023 •

edited

Loading

saunair Feb 23, 2023 •

edited

Loading