[emcai] Podcasts generated using Google's NotebookLM

[Stokoe,E]

An interesting thread about it here:

https://bsky.app/profile/carlbergstrom.com/post/3l4h45k2dlg2c 

Liz

-----Original Message-----
From: emcai <emcai-bounces at conversationanalysis.org> On Behalf Of Hendrik Buschmeier via emcai
Sent: Thursday, October 3, 2024 1:16 PM
To: L.E.M. de Rijk, (Lynn) via emcai <emcai at conversationanalysis.org>
Subject: Re: [emcai] Podcasts generated using Google's NotebookLM

Hi Roger,

No real analysis, just an impression – seconding also Magnus analysis:

The ”hosts“ communicative feedback is off and makes it unclear who knows what. In their listener behavior they produce newsmakers, as if what the other says is unknown to them, but then, in the next turn, they just take over and seemingly have all the information, and suddenly the other seems clueless. I think the model is missing that the two speakers are  independent entities with asymmetric knowledge. To me this shows how important representations of the self and the other are, even when just producing something that “just generates“ output and is not interactive.

More generally I question the merit of such applications. There are more than enough podcasts. Some people mentioned that the thing is prone to halucination as well.

From a scientific network perspective, we could think about an in depth analysis of one of these "dialogues" looking at various conversational phenomena that are produced.

Cheers

Hendrik



> On 3. Oct 2024, at 11:46, Nicolas Rollet via emcai <emcai at conversationanalysis.org> wrote:
> 
> Hi,
> 
> pauses are over-marked too, as if it's based on the fact that pauses is the key-trigger for continuers, with remind me of a very basic definition of turn in computer science, no ?
> 
> Best wishes
> 
> Nicolas Rollet
> Télécom Paris
> Department of Economics and Social Sciences
> INTERACT
> Institut polytechnique de Paris
> I3 UMR 9217 – CNRS 
> Site i3
> Encyclopédie de la parole 
> (+33)6 88 75 86 28
> bureau 3A259
> 
> De: "L.E.M. de Rijk, (Lynn) via emcai" <emcai at conversationanalysis.org>
> À: "L.E.M. de Rijk, (Lynn) via emcai" <emcai at conversationanalysis.org>
> Envoyé: Mercredi 2 Octobre 2024 18:59:11
> Objet: Re: [emcai] Podcasts generated using Google's NotebookLM
> 
> I also immediately noticed the massive overuse of continuers, acknowledgments and other response tokens. But I will say, I just tried this thing with one of our papers, and it worked very smoothly there, generating quite a realistic experience. The "podcast" really captures the majority of the contents of the paper accurately and even offers comprehensible summary descriptions of what happens in the data fragments :-O 
> 
> Quite fascinating, I must say! Thanks for sharing this, Roger.
> Best,
> Uwe
> 
> 
> Am 02.10.2024 um 18:51 schrieb Magnus Hamann via emcai:
> Just a first noticing:
> There is definitely something with the sequential distribution of "continuers" that is a little off.
> That said, this is definitely an interesting tool (thinking about students with dyslexia here).
> 
> 
> 
> Kind regards,
> Magnus Hamann
> 
> Informal office hour every Monday, Tuesday (both online), or Thursday (in my office) 11-12<Outlook-ypjhe4zs.png>
> 
> Book time to meet with me 
> <Outlook-ukk32cwg.png>From: emcai <emcai-bounces at conversationanalysis.org> on behalf of Roger K Moore via emcai <emcai at conversationanalysis.org>
> Sent: 02 October 2024 16:54
> To: emcai at conversationanalysis.org <emcai at conversationanalysis.org>
> Subject: [emcai] Podcasts generated using Google's NotebookLM
>  ** THIS MESSAGE ORIGINATED OUTSIDE LOUGHBOROUGH UNIVERSITY **
> ** Be wary of links or attachments, especially if the email is unsolicited or you don't recognise the sender's email address. **
> Hi All,
> 
> In case you are not aware, the latest AI development from Google provides the ability to generate a podcast-style conversation between two 'individuals' automatically from an uploaded document (e.g. a scientific paper)?  The results are astounding, both in terms of the speech and the dialogue.
> 
> Here are two examples I generated based on ...
>     • 
> my recent VIHAR-24 paper on "What Needs to be Known in Order to Perform a Meaningful Scientific Comparison Between Animal Communications and Human Spoken Language", and
>     • the famous "Chicken Chicken Chicken" paper by Doug Zongker.
> I wonder how these AI-generated interactions stack-up from a CA perspective?
> 
> Best wishes
> Roger
> 
> -------------------------------------------------------------------------------------------------
> Prof ROGER K MOORE* BA(Hons) MSc PhD FIOA FISCA MIET SMIEEE
> 
> Deputy Head of School
> Chair of Spoken Language Processing
> Head of Speech & Hearing Research Group (SpandH)
> Vocal Interactivity Lab (VILab), Sheffield Robotics
> School of Computer Science, UNIVERSITY OF SHEFFIELD
> Regent Court, 211 Portobello, Sheffield, S1 4DP, UK
> 
> * Winner of the 2016 Antonio Zampolli Prize for "Outstanding Contributions 
> to the Advancement of Language Resources & Language Technology 
> Evaluation within Human Language Technologies"
> 
> e-mail:  r.k.moore at sheffield.ac.uk
> web: http://staffwww.dcs.shef.ac.uk/people/R.K.Moore/
> twitter: @rogerkmoore
> Tel: +44 (0) 11422 21807
> Fax: +44 (0) 11422 21810
> Mob: +44 (0) 7910 073631
> 
> Editor-in-Chief: COMPUTER SPEECH AND LANGUAGE
> (http://www.journals.elsevier.com/computer-speech-and-language/)
> --------------------------------------------------------------------------------------------
> 
> 
> -- 
> emcai mailing list
> emcai at conversationanalysis.org
> http://conversationanalysis.org/mailman/listinfo/emcai_conversationanalysis.org
> -- 
> emcai mailing list
> emcai at conversationanalysis.org
> http://conversationanalysis.org/mailman/listinfo/emcai_conversationanalysis.org


-- 
Hendrik Buschmeier
Digital Linguistics Lab
Faculty of Linguistics and Literary Studies, Bielefeld University
https://purl.org/net/hbuschme



-- 
emcai mailing list
emcai at conversationanalysis.org
http://conversationanalysis.org/mailman/listinfo/emcai_conversationanalysis.org