Nova-2 transcriptions have the last 10-20% of the transcription cut off, often.

[venusatuluri]

Which Deepgram product are you using?

Deepgram API

Details

Nova-2 transcriptions are frequently missing the last 10-20% of the transcription. The pattern of error is very consistent - it's always the last few words that go missing. I have wav example files, but weirdly this bug report tool doesn't let me upload any standard audio formats like wav or mp3.

If you are making a request to the Deepgram API, what is the full Deepgram URL you are making a request to?

https://api.deepgram.com/v1/listen?language=en&model=nova-2&smart_format=true

If you are making a request to the Deepgram API and have a request ID, please paste it below:

No response

If possible, please attach your code or paste it into the text box.

No response

If possible, please attach an example audio file to reproduce the issue.

No response

[team-deepgram]

Thanks for asking your question about Deepgram! If you didn't already include it in your post, please be sure to add as much detail as possible so we can assist you efficiently, such as:

• The request_id if you have a question about your requests or transcription responses.
• The features you used or the full api.deepgram.com URL you sent your request to, including parameters.
• Any code snippets you can share.

[jjmaldonis]

Hey @venusatuluri we're looking into this. Can you upload the audio files you found the issue with? GitHub only allows certain file types to be uploaded so you'll have to zip the audio files before uploading them.

[venusatuluri]

nova_bug_report.zip

[jjmaldonis]

Thanks @venusatuluri very helpful. We currently think this is an issue for audio files less than 15 seconds and we're working on it. Are all your audio files very short?

[venusatuluri]

Yes, our audio files are usually less than 15 seconds.

…

On Wed, Dec 6, 2023 at 7:31 AM Jason Maldonis ***@***.***> wrote: Thanks @venusatuluri <https://github.com/venusatuluri> very helpful. We currently think this is an issue for audio files less than 15 seconds and we're working on it. Are all your audio files very short? — Reply to this email directly, view it on GitHub <#474 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABHZO2B372KRBCQH7X44WZTYICFUDAVCNFSM6AAAAABADPNGA6VHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM3TONZXG42TA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[rliffredo]

We are experiencing very similar issues.
Most of the times, however, it's on larger audio files, at least 15 seconds.

For instance, for this audio (7pk6hra.mp3.zip) we got the following response:

{
  "metadata": {
    "transaction_key": "deprecated",
    "request_id": "c043749a-2942-4647-a383-52a3559cc5f2",
    "sha256": "bd9417ff8d9c8bb0607d84b23ccafb8c72fa3cf4cba3aa26e740eab3ebce112f",
    "created": "2023-12-08T11:53:31.852Z",
    "duration": 25.8,
    "channels": 1,
    "models": [
      "a375937a-9156-40cb-940d-6a1103040ef1"
    ],
    "model_info": {
      "a375937a-9156-40cb-940d-6a1103040ef1": {
        "name": "2-general-nova",
        "version": "2023-11-14.3290",
        "arch": "nova-2"
      }
    }
  },
  "results": {
    "channels": [
      {
        "alternatives": [
          {
            "transcript": "Next time around, I wonder if I should break out the grades taught, in more detail. I think it'd be interesting to find out more about which grades in high school, perhaps even the subjects taught. So breaking out the other category into, other, you know, interspecific areas, or making sure that's the the respondents can",
            "confidence": 0.9970011,
            "words": [
              {
                "word": "next",
                "start": 0.96,
                "end": 1.28,
                "confidence": 0.99588674,
                "punctuated_word": "Next"
              },
              {
                "word": "time",
                "start": 1.28,
                "end": 1.52,
                "confidence": 0.9997961,
                "punctuated_word": "time"
              },
              {
                "word": "around",
                "start": 1.52,
                "end": 2.0,
                "confidence": 0.99558043,
                "punctuated_word": "around,"
              },
              {
                "word": "i",
                "start": 2.0,
                "end": 2.1599998,
                "confidence": 0.9997038,
                "punctuated_word": "I"
              },
              {
                "word": "wonder",
                "start": 2.1599998,
                "end": 2.56,
                "confidence": 0.9999014,
                "punctuated_word": "wonder"
              },
              {
                "word": "if",
                "start": 2.56,
                "end": 2.8,
                "confidence": 0.99989545,
                "punctuated_word": "if"
              },
              {
                "word": "i",
                "start": 2.8,
                "end": 3.04,
                "confidence": 0.99983394,
                "punctuated_word": "I"
              },
              {
                "word": "should",
                "start": 3.04,
                "end": 3.52,
                "confidence": 0.9997979,
                "punctuated_word": "should"
              },
              {
                "word": "break",
                "start": 3.52,
                "end": 3.84,
                "confidence": 0.9970011,
                "punctuated_word": "break"
              },
              {
                "word": "out",
                "start": 3.84,
                "end": 4.34,
                "confidence": 0.99784386,
                "punctuated_word": "out"
              },
              {
                "word": "the",
                "start": 4.56,
                "end": 5.04,
                "confidence": 0.999401,
                "punctuated_word": "the"
              },
              {
                "word": "grades",
                "start": 5.04,
                "end": 5.54,
                "confidence": 0.9980287,
                "punctuated_word": "grades"
              },
              {
                "word": "taught",
                "start": 5.6,
                "end": 6.1,
                "confidence": 0.885115,
                "punctuated_word": "taught,"
              },
              {
                "word": "in",
                "start": 6.72,
                "end": 7.04,
                "confidence": 0.9998895,
                "punctuated_word": "in"
              },
              {
                "word": "more",
                "start": 7.04,
                "end": 7.2799997,
                "confidence": 0.99985766,
                "punctuated_word": "more"
              },
              {
                "word": "detail",
                "start": 7.2799997,
                "end": 7.7599998,
                "confidence": 0.98655343,
                "punctuated_word": "detail."
              },
              {
                "word": "i",
                "start": 7.7599998,
                "end": 7.9199996,
                "confidence": 0.9998547,
                "punctuated_word": "I"
              },
              {
                "word": "think",
                "start": 7.9199996,
                "end": 8.08,
                "confidence": 0.99987996,
                "punctuated_word": "think"
              },
              {
                "word": "it'd",
                "start": 8.08,
                "end": 8.24,
                "confidence": 0.99879843,
                "punctuated_word": "it'd"
              },
              {
                "word": "be",
                "start": 8.24,
                "end": 8.4,
                "confidence": 0.99978024,
                "punctuated_word": "be"
              },
              {
                "word": "interesting",
                "start": 8.4,
                "end": 8.8,
                "confidence": 0.99914205,
                "punctuated_word": "interesting"
              },
              {
                "word": "to",
                "start": 8.8,
                "end": 8.96,
                "confidence": 0.9997229,
                "punctuated_word": "to"
              },
              {
                "word": "find",
                "start": 8.96,
                "end": 9.2,
                "confidence": 0.9999037,
                "punctuated_word": "find"
              },
              {
                "word": "out",
                "start": 9.2,
                "end": 9.36,
                "confidence": 0.9998086,
                "punctuated_word": "out"
              },
              {
                "word": "more",
                "start": 9.36,
                "end": 9.86,
                "confidence": 0.9994623,
                "punctuated_word": "more"
              },
              {
                "word": "about",
                "start": 10.304999,
                "end": 10.705,
                "confidence": 0.9397779,
                "punctuated_word": "about"
              },
              {
                "word": "which",
                "start": 10.705,
                "end": 11.184999,
                "confidence": 0.999833,
                "punctuated_word": "which"
              },
              {
                "word": "grades",
                "start": 11.184999,
                "end": 11.584999,
                "confidence": 0.9696138,
                "punctuated_word": "grades"
              },
              {
                "word": "in",
                "start": 11.584999,
                "end": 11.745,
                "confidence": 0.9996402,
                "punctuated_word": "in"
              },
              {
                "word": "high",
                "start": 11.745,
                "end": 11.905,
                "confidence": 0.9995679,
                "punctuated_word": "high"
              },
              {
                "word": "school",
                "start": 11.905,
                "end": 12.405,
                "confidence": 0.96477866,
                "punctuated_word": "school,"
              },
              {
                "word": "perhaps",
                "start": 13.025,
                "end": 13.424999,
                "confidence": 0.58096147,
                "punctuated_word": "perhaps"
              },
              {
                "word": "even",
                "start": 13.424999,
                "end": 13.664999,
                "confidence": 0.99357784,
                "punctuated_word": "even"
              },
              {
                "word": "the",
                "start": 13.664999,
                "end": 13.825,
                "confidence": 0.9576805,
                "punctuated_word": "the"
              },
              {
                "word": "subjects",
                "start": 13.825,
                "end": 14.304999,
                "confidence": 0.99692434,
                "punctuated_word": "subjects"
              },
              {
                "word": "taught",
                "start": 14.304999,
                "end": 14.705,
                "confidence": 0.9293262,
                "punctuated_word": "taught."
              },
              {
                "word": "so",
                "start": 14.705,
                "end": 15.205,
                "confidence": 0.9992028,
                "punctuated_word": "So"
              },
              {
                "word": "breaking",
                "start": 15.344999,
                "end": 15.664999,
                "confidence": 0.8038731,
                "punctuated_word": "breaking"
              },
              {
                "word": "out",
                "start": 15.664999,
                "end": 15.905,
                "confidence": 0.9981931,
                "punctuated_word": "out"
              },
              {
                "word": "the",
                "start": 15.905,
                "end": 16.145,
                "confidence": 0.72913694,
                "punctuated_word": "the"
              },
              {
                "word": "other",
                "start": 16.145,
                "end": 16.465,
                "confidence": 0.9968208,
                "punctuated_word": "other"
              },
              {
                "word": "category",
                "start": 16.465,
                "end": 16.965,
                "confidence": 0.998252,
                "punctuated_word": "category"
              },
              {
                "word": "into",
                "start": 17.025,
                "end": 17.525,
                "confidence": 0.7878603,
                "punctuated_word": "into,"
              },
              {
                "word": "other",
                "start": 18.32,
                "end": 18.72,
                "confidence": 0.8469903,
                "punctuated_word": "other,"
              },
              {
                "word": "you",
                "start": 18.72,
                "end": 18.880001,
                "confidence": 0.9902318,
                "punctuated_word": "you"
              },
              {
                "word": "know",
                "start": 18.880001,
                "end": 19.12,
                "confidence": 0.99799347,
                "punctuated_word": "know,"
              },
              {
                "word": "interspecific",
                "start": 19.12,
                "end": 19.62,
                "confidence": 0.931517,
                "punctuated_word": "interspecific"
              },
              {
                "word": "areas",
                "start": 19.92,
                "end": 20.42,
                "confidence": 0.63309896,
                "punctuated_word": "areas,"
              },
              {
                "word": "or",
                "start": 21.04,
                "end": 21.12,
                "confidence": 0.5008061,
                "punctuated_word": "or"
              },
              {
                "word": "making",
                "start": 21.12,
                "end": 21.44,
                "confidence": 0.9887378,
                "punctuated_word": "making"
              },
              {
                "word": "sure",
                "start": 21.44,
                "end": 21.6,
                "confidence": 0.99133146,
                "punctuated_word": "sure"
              },
              {
                "word": "that's",
                "start": 21.6,
                "end": 21.92,
                "confidence": 0.6207464,
                "punctuated_word": "that's"
              },
              {
                "word": "the",
                "start": 21.92,
                "end": 22.42,
                "confidence": 0.50518316,
                "punctuated_word": "the"
              },
              {
                "word": "the",
                "start": 22.48,
                "end": 22.56,
                "confidence": 0.6938833,
                "punctuated_word": "the"
              },
              {
                "word": "respondents",
                "start": 22.56,
                "end": 23.06,
                "confidence": 0.988893,
                "punctuated_word": "respondents"
              },
              {
                "word": "can",
                "start": 24.16,
                "end": 24.66,
                "confidence": 0.8660644,
                "punctuated_word": "can"
              }
            ],
            "paragraphs": {
              "transcript": "Next time around, I wonder if I should break out the grades taught, in more detail. I think it'd be interesting to find out more about which grades in high school, perhaps even the subjects taught. So breaking out the other category into, other, you know, interspecific areas, or making sure that's the the respondents can",
              "paragraphs": [
                {
                  "sentences": [
                    {
                      "text": "Next time around, I wonder if I should break out the grades taught, in more detail.",
                      "start": 0.96,
                      "end": 7.7599998
                    },
                    {
                      "text": "I think it'd be interesting to find out more about which grades in high school, perhaps even the subjects taught.",
                      "start": 7.7599998,
                      "end": 14.705
                    },
                    {
                      "text": "So breaking out the other category into, other, you know, interspecific areas, or making sure that's the the respondents can",
                      "start": 14.705,
                      "end": 24.66
                    }
                  ],
                  "num_words": 56,
                  "start": 0.96,
                  "end": 24.66
                }
              ]
            }
          }
        ]
      }
    ]
  }
}

[jjmaldonis]

Thanks @rliffredo, we are currently looking into this.

[danielclas]

Hi @jjmaldonis having the same issue as above with nova-2, using Node SDK. Any news on that?

[jjmaldonis]

Hey all (@venusatuluri @rliffredo @danielclas), we released an update to our Nova 2 model today that prevents missing words at the end of transcripts. Now, when you use model=nova-2-general or model=nova-2, you should not find missing words at the end of transcripts. If you test, please let us know your results.

We are still updating the Nova 2 phonecall, meeting, etc. models with this fix and the updated models will be released as they are updated.

[mattbonnell]

has the phonecall model been fixed?

the updated models will be released as they are updated

is there a release page or something where I can track this?

[jjmaldonis]

Hey @mattbonnell, we're still working on the phonecall model. We don't have a public way to track this info right now but feel free to keep asking.

We released updates to our nova-2-general streaming models earlier this week.

[retry-the-user]

We are seeing a similar problem very intermittently while using nova-2 phonecall model. I have just reverted us back to the nova phonecall model. Sometimes text would be missing from the start and not just a word or two, but an entire sentence. In another example, it was missing a bunch of text from the middle, which makes even less sense.

If you want to look at the issue here is log data for you:
SyZ81BPf.wav 65e2d92f-69ae-4d52-8b7f-912e4bdcacac size: 1700894 stt_req: 0.35 2024-01-25 19:33:17

This is a 1m30s piece of audio. The whole first 15 seconds of audio is not transcribed. This is what we got back from you:

"We can work up a no obligation quote for you. Does that sound good? Sure. I apologize for not providing that information upfront. Our company name is Tribeca Group located in New York. Our callback number is x x x x x x, and our website is ww.tribeccagroup.com. Is there anything else I can assist you with? Absolutely. I apologize for not providing that information earlier. Our company name is Tribeca Group located in New York. You can reach us at x x x x x, and our website is ww.tribeccagroup.com. Is there anything else I can help you with? I apologize for any confusion. Tribeca Group is a reputable company based in New York. We specialize in providing capital for businesses like yours. Our callback number is x x x x x x, and you can find more information about us on our website, wwtribeccagroup.com. We're here to assist you with any questions or concern"

[jjmaldonis]

Hey @retry-the-user, we released an updated version of nova-2-phonecall last Friday with a fix. I just tested the new vs. old model: I can reproduce the missing sentence on the old model, and the new model correctly transcribes all words/sentences. You should be good to go now!

[venusatuluri]

feb5_update.zip
Thank you for the updates, my original examples are all working correctly now.

However, I ran into the same or similar issue with a different file today - the transcription is empty with nova-2, while nova returns the transcription correctly. It's a rather short audio, for sure. I have attached the zip file.

[venusatuluri]

feb6_update.zip
Adding another example where the ending is again cut off, and the issue happens in nova-2 but not in nova. I've observed the same issue with the corresponding phonecall models.

[jjmaldonis]

Hey @venusatuluri, thanks for these examples and the run_and_compare.sh scripts. I will pass these to our team.

[rwrz]

It just happened to me as well. We are transcribing big podcast episodes and some sentences are missing!

Request ID: f85f6f12-8e8d-493e-aa3f-584f01accdad
Performed at: 02/10/2024 10:28:49

File: https://storage.googleapis.com/test-reallife-app/rev/RLEP%20378%20(Transcript%20v3).flac.zip

Missing sentences (you can see the "paragraph" timestamp on the top-left corner of each picture):

Having a better model like Nova-2, which reduces the WER is amazing. But if it keeps missing a few sentences, it will increase our "review processes" time much more than having some words wrong. Not worthing migrating.

[jjmaldonis]

Hey @rwrz, I wasn't able to reproduce the missing words in this podcast. Do you have a JSON file containing DG's output for the request you made?

[rwrz]

Sure, here is the json file.
BTW, I'm using these parameters:

transcript_377_deepgram.json.zip

[rwrz]

Updates on this?

[rwrz]

I've tried more than once and every time it has the very same missing sentences. I believe this is an algorithm issue with the formatting, multichannel, or similar. Should be easy to spot and fix.

[vpalenik1]

Hello @team-deepgram We are also experiencing issues with missing text in the transcripts. We noticed that recordings with a bit lower audio quality are missing transcription at the beginning and also at the end. This issue happens only using the Nova 2 model. We transcribed the same audio with the Deepgram's Whisper model and no text was missing. Such issues make the Nova 2 model totally unreliable. The issue impacts not only English but also other languages. We need a permanent solution as it seems to be a problem for many of your customers.

[vpalenik1]

@team-deepgram @jjmaldonis It looks like many of your customers are experiencing the same issue. Can you please give us an update? Is Deepgram working on the fixes? Thank you.

[vpalenik1]

@jkroll-deepgram

[NorthbridgeBB]

What is the status on this concerning issue, we are now at a decision point on Deepgram or Whisper 3

[vpalenik1]

I haven't heard back from anyone from Deepgram for more than 3 weeks! @team-deepgram can you please provide some update?

[NorthbridgeBB]

Thanks for the heads up, Deepgram scratched, poor support.

[jkroll-deepgram]

Hi all who are on this thread (cc @NorthbridgeBB @vpalenik1 @rwrz).

My apologies that our models are not meeting your expectations.

Models have their relative strengths and weaknesses depending on various aspects of audio - such as your audio length, format, quality, background noise, streaming versus pre-recorded audio, and also your business use case.

We offer several different models and encourage users to do their own testing to see what works best for them. For customers such as yourselves who don't see results that they're satisfied with using Nova-2 (model=nova-2), we do also have a previous Nova model (model=nova), and our prior Enhanced architecture (model=enhanced).

I have also seen that file format specifically makes a difference with pre-recorded audio for Nova-2, where converting mp3 to wav reduces missed words.

For further suggestions on how you can test different conditions with your own audio, also see this thread: https://github.com/orgs/deepgram/discussions/623

While some patterns of errors may appear similar on the surface, often there are various underlying factors due to ML model training, the audio itself, and the application needs, rather than being a single bug that can be trivially located and fixed.

We are continually iterating and training new models in response to customer feedback as well as innovating on our model architectures.

I understand that Nova-2 does have cases where it is "shyer" and may miss words in a transcript. We have researched and improved several known causes of this pattern of results, and based on your feedback, we have more work to continue to do.

While we may not be able to triage and root-cause each unsatisfactory request, I invite you to share Deepgram request IDs, which is the most effective way for us to internally examine and investigate any poorly-performing requests in our ongoing efforts to improve our models.

[rwrz]

I've experimented a few things, turning on and off stuff, and I hope it can help anyone that is experiencing these issues as well.
You can experiment yourself, using this audio.

Multichannel Issue [multichannel]

During my tests, this option will increase the probability to have fully missing sentences.
You can replace this with DIARIZATION instead, fixes 90% of the issues.

With multichannel:

"It's kinda shocking that when she tries to communicate Barcelona."

With DIARIZATION instead [docs]:

"It's kinda shocking that when she tries to communicate in Spanish in different situations when, for example, they go to a cafe, a coffee shop in Barcelona."

SmartFormat issue [docs]

During my tests, I've experienced some sentences repeating, sometimes even "unexpected words instead". You can replace it with PUNCTUATION if you need to, it works much better. Yeah, it is not the same, but produces more reliable outputs.

With Smart Format:

"And that means a lot because it is reinforcing identity. So it''s a virtuous cycle.\n\nWow. The identity. So it''s a virtuous cycle."

Correct (with punctuation) [docs]:

And that means a lot because it is reinforcing identity is gonna lead you to having more solid habits, which will then reinforce back the identity. So it's a virtuous cycle.

--
I hope this report would help @team-deepgram (@jkroll-deepgram) to find those issues. They all seems to be related to their "formatting" algorithms, not related to the models at all.

[venusatuluri]

Hi, I tried nova-2 after a few months today to see if this problem has resolved. Unfortunately the pattern persists where nova returns the correct transcription while nova-2 returns empty string.

The request id using nova-2 is 30f7ce1c-bc91-411e-b7d2-7ae7814d04ca, while the request id using nova is 4fa00dd3-a50b-4ca1-b831-d37467d4db8f.

[rwrz]

The problems I was facing, they were better for a while, but today we had a big failure on this sample:
output_0_6.zip

It is crazy how it can halucinate like this.
Sentences with the RED lines are wrong OR has wrong timestamps.

I hope that having a sample of the issue, will help you to fix it.