[HN Gopher] Trying out Gemini 3 Pro with audio transcription and...
       ___________________________________________________________________
        
       Trying out Gemini 3 Pro with audio transcription and a new pelican
       benchmark
        
       Author : nabla9
       Score  : 90 points
       Date   : 2025-11-18 19:05 UTC (3 hours ago)
        
 (HTM) web link (simonwillison.net)
 (TXT) w3m dump (simonwillison.net)
        
       | simonw wrote:
       | The audio transcript exercise here is particularly interesting
       | from a journalism perspective.
       | 
       | Summarizing a 3.5 hour council meeting is something of a holy
       | grail of AI-assisted reporting. There are a LOT of meetings like
       | that, and newspapers (especially smaller ones) can no longer
       | afford to have a human reporter sit through them all.
       | 
       | I tried this prompt (against audio from
       | https://www.youtube.com/watch?v=qgJ7x7R6gy0):
       | Output a Markdown transcript of this meeting. Include speaker
       | names and timestamps. Start with an outline of the key
       | meeting sections, each with a title and summary and timestamp
       | and list of participating names. Note in bold if anyone
       | raised their voices, interrupted each other or had
       | disagreements. Then follow with the full transcript.
       | 
       | Here's the result:
       | https://gist.github.com/simonw/0b7bc23adb6698f376aebfd700943...
       | 
       | I'm not sure quite how to grade it here, especially since I
       | haven't sat through the whole 3.5 hour meeting video myself.
       | 
       | It appears to have captured the _gist_ of the meeting very well,
       | but the fact that the transcript isn 't close to an exact match
       | to what was said - and the timestamps are incorrect - means it's
       | very hard to trust the output. Could it have hallucinated things
       | that didn't happen? Those can at least be spotted by digging into
       | the video (or the YouTube transcript) to check that they
       | occurred... but what about if there was a key point that Gemini 3
       | omitted entirely?
        
         | WesleyLivesay wrote:
         | I think it appears to have done a good job of summarizing the
         | points that it summarize, at least judging from my quick watch
         | of a few sections and from the YT Transcript (which seems quite
         | accurate).
         | 
         | Almost makes me wonder if it is behind the scenes doing
         | something similar to: rough transcript -> Summaries ->
         | transcript with timecodes (runs out of context) -> throws
         | timestamps that it has on summaries.
         | 
         | I would be very curious to see if it does better on something
         | like an hour long chunk of audio, to see if it is just some
         | sort of context issue. Or if this same audio was fed to it in
         | say 45 minute chunks to see if the timestamps fix themselves.
        
         | byt3bl33d3r wrote:
         | I've been meaning to create & publish a structured extraction
         | benchmark for a while. Using LLMs to extract
         | info/entities/connections from large amounts of unstructured
         | data is also a huge boon to AI-assisted reporting and has also
         | a number of cybersecurity applications. Gemini 2.5 was pretty
         | good but so far I have yet to see an LLM that can reliably ,
         | accurately and consistently do this
        
           | simonw wrote:
           | This would be extremely useful. I think this is one of the
           | most commercially valuable uses of these kinds of models,
           | having more solid independent benchmarks would be great.
        
         | mistercheph wrote:
         | For this use case I think best bet is still a toolchain with a
         | transcription model like whisper fed into an LLM to summarize
        
           | simonw wrote:
           | Yeah I agree. I ran Whisper (via MacWhisper) on the same
           | video and got back accurate timestamps.
           | 
           | The big benefit of Gemini for this is that it appears to do a
           | great job of speaker recognition, plus it can identify when
           | people interrupt each other or raise their voices.
           | 
           | The best solution would likely include a mixture of both -
           | Gemini for the speaker identification and tone-of-voice
           | stuff, Whisper or NVIDIA Parakeet or similar for the
           | transcription with timestamps.
        
         | rahimnathwani wrote:
         | For this use case, why not use Whisper to transcribe the audio,
         | and then an LLM to do a second step (summarization or answering
         | questions or whatever)?
         | 
         | If you need diarization, you can use something like
         | https://github.com/m-bain/whisperX
        
           | pants2 wrote:
           | Whisper simply isn't very good compared to LLM audio
           | transcription like gpt-4o-transcribe. If Gemini 3 is even
           | better it's a game-changer.
        
           | crazysim wrote:
           | Since Gemini seems to be sucking at timestamps, perhaps
           | Whisper can be used to help ground that as an additional
           | input alongside the audio.
        
         | ks2048 wrote:
         | Does anyone benchmark these models for text-to-speech using
         | traditional word-error-rates? It seems audio-input Gemini is a
         | lot cheaper than Google Speech-to-text.
        
           | simonw wrote:
           | Here's one: https://voicewriter.io/speech-recognition-
           | leaderboard
           | 
           | "Real-World Speech-to-text API Leaderboard" - it includes
           | scores for Gemini 2.5 Pro and Flash.
        
         | Workaccount2 wrote:
         | My assumption is that Gemini has no insight into the time
         | stamps, and instead is ballparking it based on how much context
         | has been analyzed up to that point.
         | 
         | I wonder if you put the audio into a video that is nothing but
         | a black screen with a timer running, it would be able to
         | correctly timestamp.
        
           | simonw wrote:
           | The Gemini documentation specifically mentions timestamp
           | awareness here: https://ai.google.dev/gemini-api/docs/audio
        
           | minimaxir wrote:
           | Per the docs, Gemini represents each second of audio as 32
           | tokens. Since it's a consistent amount, as long as the model
           | is trained to understand the relation between timestamps and
           | the number of tokens (which per Simon's link it does), it
           | should be able to infer the correct amount of seconds.
        
         | potatolicious wrote:
         | You really want to break a task like this down to constituent
         | parts - especially because in this case the "end to end" way of
         | doing it (i.e., raw audio to summary) doesn't actually get you
         | anything.
         | 
         | IMO the right way to do this is to feed the audio into a
         | transcription model, specifically one that supports diarization
         | (separation of multiple speakers). This will give you a high
         | quality raw transcript that _is_ pretty much exactly what was
         | actually said.
         | 
         | It would be rough in places (i.e., Speaker 1, Speaker 2, etc.
         | rather than actual speaker names)
         | 
         | Then you want to post-process with a LLM to re-annotate the
         | transcript and clean it up (e.g., replace "Speaker 1" with
         | "Mayor Bob"), and query against it.
         | 
         | I see another post here complaining that direct-to-LLM beats a
         | transcription model like Whisper - I would challenge that. Any
         | modern ASR model will do a very, very good job with 95%+
         | accuracy.
        
           | simonw wrote:
           | Which diarization models would you recommend, especially for
           | running on macOS?
           | 
           | (Update: I just updated MacWhisper and it can now run
           | Parakeet which appears to have decent diarization built in,
           | screenshot here:
           | https://static.simonwillison.net/static/2025/macwhisper-
           | para... )
        
           | darkwater wrote:
           | Why can't Gemini, the product, do that by itself? Isn't the
           | point of all this AI hype to easily automate things with low
           | effort?
        
             | vlovich123 wrote:
             | Multimodal models are only now starting to come into the
             | space and even then I don't know they really support
             | diarization yet (and often multimodal is
             | thinking+speech/images, not sure about audio).
        
           | sillyfluke wrote:
           | I'm curious when we started conflating transcription and
           | summarization when discussing this LLM mess, or maybe I'm
           | confused about the output simonw is quoting as "the
           | transcript" which starts off not with the actual transcript
           | but with a Meeting Outline and Summarization sections?
           | 
           | LLM summarization is utterly useless when you want 100%
           | accuracy on the final binding decisions on things like
           | council meeting decisions. My experience has been that LLMs
           | cannot be trusted to follow convulted discussions, including
           | revisting earlier agenda items later in the meeting etc.
           | 
           | With transcriptions, the catastrophic risk is far less since
           | I'm doing the summarizing from a transcript myself. But in
           | that case, for an auto-generated transcript, I'll take
           | correct timestamps with gibberish sounding sentences over
           | incorrect timestamps with "convincing" sounding but
           | halluncinated sentences any day.
           | 
           | Any LLM summarization of a sufficiently important meeting
           | requires second-by-second human verification of the audio
           | recording. I have yet to see this convincingly refuted (ie,
           | an LLM model that maintains 100% accuracy on summarizing
           | meeting decisions consistently).
        
             | simonw wrote:
             | That's why I shared these results. Understanding the
             | difference between LLM summarization and exact
             | transcriptions is really important for this kind of
             | activity.
        
             | Royce-CMR wrote:
             | This is a high area of focus for me and I agree: following
             | a complex convo, especially when it gets picked up again
             | 20-30 min later, is difficult.
             | 
             | But not impossible. I've had success with prompts that ID
             | all topics and then map all conversation tied to each topic
             | (each seperate LLM queries) and then pulling together
             | summary and conclusions by topic.
             | 
             | I've also had success with one shot prompts - especially
             | with the right context on the event and phrasing shared.
             | But honestly I end up spending about 5-10 min reviewing and
             | cleaning up the output before solid.
             | 
             | But that's worlds better than attending the event, and then
             | manually pulling together notes from your fast in flight
             | shorthand.
             | 
             | (Former BA, ran JADs etc, lived and died by accuracy and
             | right color / expression / context in notes)
        
       | londons_explore wrote:
       | Anyone got a class full of students and able to get a human
       | version of this pelican benchmark?
       | 
       | Perhaps half with a web browser to view the results, and half
       | working blind with the numbers alone?
        
       | ZeroConcerns wrote:
       | > so I shrunk the file down to a more manageable 38MB using
       | ffmpeg
       | 
       | Without having an LLM figure out the required command line
       | parameters? Mad props!
        
         | simonw wrote:
         | Hah, nope! I had Claude Code figure that one out.
        
       | leetharris wrote:
       | I used to work in ASR. Due to the nature of current multimodal
       | architectures, it is unlikely we'll ever see accurate timestamps
       | over a longer horizon. You're better off using encoder-decoder
       | ASR architectures, then using traditional diarization using
       | embedding clustering, then using a multimodal model to refine it,
       | then use a forced alignment technique (maybe even something pre-
       | NN) to get proper timestamps and reconciling it at the end.
       | 
       | These things are getting really good at just regular
       | transcription (as long as you don't care about verbatimicity),
       | but every additional dimension you add (timestamps, speaker
       | assignment, etc) will make the others worse. These work much
       | better as independent processes that then get reconciled and
       | refined by a multimodal LLM.
        
       | nurumaik wrote:
       | Seems like pelican benchmark is finally added to model training
       | process
        
       | Wowfunhappy wrote:
       | Aww, I don't like the new pelican benchmark as much. I liked that
       | the old prompt was vague and we could see how the AI interpreted
       | it.
        
         | ahmedfromtunis wrote:
         | Yeah. The new challenge seems easier to solve since it
         | basically is hand-holding the LLMs into what the result should
         | look like.
         | 
         | I think a more challenging, well, challenge, would be to offer
         | an even more absurd scenario and see how the model handles it.
         | 
         | Example: generate an svg of a pelican and a mongoose eating
         | popcorn inside a pyramid-shaped vehicle flying around Jupiter.
         | Result: https://imgur.com/a/TBGYChc
        
           | simonw wrote:
           | I like the hand-holding because it's a better test of how
           | well models can follow more detailed instructions.
           | 
           | I was inspired by Max Woolf's nano banana test prompts:
           | https://minimaxir.com/2025/11/nano-banana-prompts/
        
             | ahmedfromtunis wrote:
             | That's a valid point but I'd argue the new test would be
             | then interesting to couple with the original one, not to
             | replace it.
             | 
             | Do you think it would be reasonable to include both in
             | future reviews, at least for the sake of back-compatibility
             | (and comparability)?
        
               | simonw wrote:
               | Yeah I'm going to keep on using the old one as well.
        
       | razodactyl wrote:
       | I was waiting for this post.
       | 
       | Love the pivot in pelican generation bench.
        
       ___________________________________________________________________
       (page generated 2025-11-18 23:00 UTC)