Making 1,000 Audio Lectures Accessible: 92NY, Sesling Precision Group, and TheirStory:

March 4, 2026 - Zack Ellis

Participants

Michael Sesling, Hanassa Wicks, Caitlin Biggers, SPEAKER_S4, SPEAKER_S5, SPEAKER_S6, SPEAKER_S7, SPEAKER_S8

Summary

The conversation primarily centered on a collaborative project involving the transcription and authentication of archival recordings from the 92nd Street Y's Unterberg Poetry Center. This initiative, supported by the Mellon Foundation, aimed to make the Center's extensive collection of literary event recordings more accessible. The participants discussed the integration of AI technologies, specifically TheirStory, to enhance the transcripts' accuracy while managing costs. Challenges such as handling problematic audio, non-English languages, and ensuring high transcription quality were addressed through a combination of AI and human authentication. The project resulted in a publicly accessible, searchable archive with innovative tools that identify themes across recordings, thereby enabling a broader audience to explore the collection efficiently.

0:00 Michael Sesling

Forward. Uh, precision Group is the new firm coming out of it? Uh, I'm going to hand it off to Hanassa Wicks to introduce herself.

0:11 Hanassa Wicks

Good morning. Well, technically, it's still morning for me. I'm in central time, so good morning. Good afternoon. Wherever you're at. Um, I'm. I was a part of Michael's team with ATC and have since migrated. Still as a part of Michael's team. Um, I do have a small business, uh, virtual administrative business and project management called Wicks Management. Um, I am located in the Up of Michigan, and those are just kind of some of the things that I do. I'm really sorry. I really do an array of things. I will hand it off to you.

0:52 Caitlin Biggers

Hi, everyone. I'm Caitlin Biggers. Um, I am the archivist at the 92nd Street Y, and we are on the Upper East Side of Manhattan. Um, I've worked here since 2023. And basically, I oversee the organization's, uh, large archive. And my primary goal is to get as much of it in the hands of the public as possible. Um, so I was fortunate to jump into this transcription and authentication project right away with these guys.

1:21 SPEAKER_S4

Thank you. Caitlin. Hi, everybody. I am Zack Ellis, the founder and CEO of TheirStory. Uh, we are purpose built platform for oral history and audiovisual knowledge and story management. Uh, grew up in Rochester, New York, spent some time in Silicon Valley, uh, before moving back to Rochester. And as of last week, I am now a new resident of Brooklyn, New York. So, um, very happy to be in New York City, uh, along with Caitlin as well. Um, in terms of TheirStory, we're used by over 120 organizations, uh, around both the United States, Canada and Singapore and Ghana.

2:00 SPEAKER_S4

So, uh, international, uh, in terms of usage, which is very exciting and, um, happy to be here and to have been in support of this project over the last couple of years. I'll pass it over to Michael.

2:14 Michael Sesling

Thank you, Zack, and congratulations on the move to Brooklyn. Zack.

2:18 SPEAKER_S4

Thank you very much.

2:19 Michael Sesling

Um, partnerships have been really important as we work together. And Zack and I first met in 2020. Holy moly. Six years ago, Zack and I started talking about ways that we could potentially work together as technology and humans. And we had, uh, at the previous firm previously worked with the 92nd Street Y, as you see on the bottom right, the Elie Wiesel Archive, where we worked on 180 recordings with humans, transcribing, uh, 270 hours for that project and authenticating our own work as well. Authentication. Meaning we were after we completed transcribing the project, listening along, reading along, and making any corrections along the way. The archive was 50 plus years of Elie Wiesel's speeches and readings that he gave at the 92nd Street Y, with some absolutely phenomenal content. After we had finished that project, the 92nd Street Y had reached back out to me about this new project where they had 1000 recordings, approximately 1500 hours of content that they wanted to have transcribed. And when we looked at that, and we knew the cost of such a large scale project with humans.

3:50 SPEAKER_S5

Or an investigator, you know.

3:54 Michael Sesling

Well, thank you. If everybody can just check that they are muted I would appreciate it. Thank you. Um. And we went from there, um, reaching out to Zack at TheirStory to figure out a way that we could utilize technology to help process all of those recordings and help cut down on those costs, but still utilize humans to authenticate the AI output. Um, and the three organizations started Kibitzing, as one might say, and looking at ways that we could all work together to make this process work.

4:36 Caitlin Biggers

So just to provide a little bit of info on the material that we were working with and a little background on the why itself, um, the 92nd Street Y is a cultural and community center in New York City. Uh, we have over 150 years and counting, um, of history, supporting arts and cultural programming, amongst many other things like we have a gym, people live here. It's it's also why, in the traditional sense, um, one of our, uh, most prominent, uh, centers is the Unterberg Poetry Center. It's one of our unique programming programming areas. And, uh, we hold an incredible collection of event recordings dating back to 1949. These recordings feature the biggest names in in literature, uh, past and present. We have T.S. Eliot. We have premieres from Truman Capote, Kurt Vonnegut, Susan Sontag, Joan Didion. Uh, really, anyone you can think of in the 20th century. Um, and today it's it's not out of the question that they've spoken at the Y. So it's a really, really special collection. Um, all of these event recordings were made here at the Y, and they typically feature authors reading selections of their work, uh, sometimes solo, sometimes in conversation with another performer or another introducer who's also typically someone who's pretty keyed into the poetry center, um, or the poetry world. Um, if you could hit the slide. Zack. Um, so caring for this collection was a unique challenge because I was joining not only an organization that I had no history with, but there was also an active grant project that was underway as well. Um, Michael gave a little bit of background, but I soon learned that there had actually been, um, talks of doing a project like this for nearly three years before I joined the team, and the grant actually needed additional revisions in order to move forward. So just to take a step back a bit, um, the whole project began in 2021 as a way to use innovative archive work to create groundbreaking scholarship and preserve our incredible collection, and also to invite new audiences into the collection who might not know that we even hold this material. Um, we had been working with the Mellon Foundation to refine a proposal that achieve these goals and also met the objectives of the foundation as well. We knew we wanted to transcribe the recordings because transcription really makes up all of the recordings instantly easier to access, especially when we're talking about audio. Um, and this is especially important for us because most of these recordings are an hour and a half long and not everyone has, you know, the time to sit down and listen to an hour and a half long recording. Sometimes they just want to hear the reading of this specific poem, something like that. Um, quickly. As Michael mentioned, it became clear that time and consequently budget would be the biggest factors in shaping our project. So we knew how long it took to have humans transcribe recordings from our previous experience working on the project with ATC. Um, and as he mentioned, that that covered about 270 hours of audio, uh, depending on the version that we were considering of this project, we were looking at anywhere from 1500 to potentially 3000 hours of audio. So right from the the start, we knew that AI would need to be a part of this project. Um, of course, still accuracy was very important because these are poets reading from their work. We cannot have AI putting words in their mouth or getting things wrong. That would fundamentally change the nature of what was presented on stage. So transcript. Authentication was a part of the conversation at nine, two and one right away. Ultimately, we revised our proposal more than five times, um, iterating through different recording totals, different value adding elements, different inclusions, um, before landing on the completed version, uh. Throughout this process, the Mellon Foundation made recommendations that help us not only strengthen our proposal, but also our project. So we're we're grateful to them. They're, um, still, after all of this, arriving at a project that we could actually embark on and really take off on was no small feat.

8:47 Caitlin Biggers

Ultimately, we decided to transcribe and authenticate 1000 recordings, which works out to about 1500 hours. And we also applied some additional value adding tools that I will discuss in a few. This one's Michael, I believe. Oh.

9:09 Michael Sesling

I forgot I was muted. My apologies. Um, the challenges that we found around, uh, grants and working with the 92nd Street Y and full disclosure, it was Caitlin's predecessor that was working on the grant. Um, and it was one of the challenges that we had in partnering once Caitlin took over, was trying to piece together, um, all that had happened on the 92nd Street y side of things prior to Caitlin's arrival. Um, but we did learn that humans, as I mentioned earlier, were too expensive for 1500 hours of Recordings, um, for the grant that the 92nd Street Y was looking to receive. Key factors were that they also wanted to find a way to ensure that, uh, this work would still uphold 99% accuracy, uh, with the transcription, which we knew that working with AI alone would not offer that. And that's when we started looking at the authentication and having, uh, my team work in conjunction with TheirStory to process all of those recordings, uh, which we will get into, uh, in a few moments. Um. Oh, sorry about that, Zack. Um. When we looked at all of the variabilities, um, in this process and we looked at, um, what are the costs that we can. Pay our team to do this work where all of the people doing the work as transcriptionists are still earning the dollars that they deserve to be paid. And no matter how we crunch those numbers, it was impossible to make it work. Um. With, uh, the partnership with TheirStory, um, we found a way that we were able to make those figures work and allow for the humans on my team to still be, uh, pardon me, to ultimately be trained as authentication specialists, as we call them, uh, reviewing the work of AI output. Now, you can go ahead, Zack.

11:37 SPEAKER_S4

Thank you. Michael. So it was an interesting time for us to come into the project. Um. Ultimately, when Michael had first reached out to me and then we spoke with, uh, then head of the Poetry center, uh, at nine two and y um, the initial idea was actually to use an existing

12:00 SPEAKER_S4

transcription platform that, uh, that the Y had had access to, um, in com uh, but there were a few reasons that as we all, uh, talked about, that it started to make sense for, uh, for TheirStory to be a part of the project. So one was that, um, as we thought about how to make the collection most usable for researchers and for the public, it became clear that the transcript itself wasn't the only output that we wanted to make available. Um, it was also, uh, creating indexes or chapter chapter structures, um, so someone could quickly read summaries, um, at key moments throughout any recording just to make it easier for for the researcher. Um, it also was a time when, uh, new transcription technologies were just coming out, uh, a main one being whisper, uh, the open source, uh, speech to text model that was released by OpenAI. And one of the jobs that we have at TheirStory is really to build or integrate the best of what is out there into the platform to make it available for other institutions that might not be building their own software. Uh, and so this was a key change in an improvement in the accuracy of speech to text. Whereas before this, uh, it would actually take as much time, if not more sometimes for humans to correct AI generated transcripts than it was to actually just transcribe from scratch. And so this was a key innovation that we were able to bring into TheirStory to bear, uh, at this project. Otherwise it just, again, would not have been possible to, uh, to do, uh, with the combination of AI plus plus humans. Um, this was also something where we knew that needs may change over time. And so having a platform that was purpose built for, um, for this process to go from ingest to processing the recordings to curating the recordings publicly, um, which is also something that, um, that we had been building out during this time. Um, it was going to be important to have, uh, a partner, uh, that could adapt with whatever the changing needs of the, of the project were. Um, and so fortunately, Michael and I have been building relationship over several years and had that trust to be able to, uh, to move together and to adapt to whatever the needs were of 92nd Street y of melon. Um, and I think that actually proved to be really important in the beginning stages when, uh, after we lost the, uh, the, the director and had to shift, um, our stance, uh, we were able to work on the idea of actually, instead of just having a couple Researchers do original scholarship to build a research portal that would allow any researcher to take advantage of, uh, of the collection. And we can talk a little bit more about that and what it looks like and, and actually how we're now making that available to, to anyone as an open source infrastructure. So we're very grateful for this project and collaboration to have made both the infrastructure itself open source for anyone to create a similar research portal, but also some of the other innovations we had during this project, like creating a custom machine learning model to extract named entities so people, places, events, dates,

15:20 SPEAKER_S4

literary works of art, um, all again making it easier for researchers to to make use, make sense of and explore, uh, the collection, uh, one of the last pieces that we didn't expect when we first started was actually being able to embed, uh, an accessible audio visual player on the 92nd Street Wise website itself for the over 850 recordings that are now public. So all these things were, uh, what, um, made it a good fit, uh, where we were able to adapt, uh, over time to, to the needs of the project. I'll pass it over to, Hanassa, to talk about, uh, uh, kind of what the process looked like, um, uh, to actually manage a thousand recordings that we were ultimately processing because AI can cannot do it alone.

16:14 Hanassa Wicks

All right. Thank you. Um, so, yes, like Zack said, if we left it up to AI alone and didn't have humans touch the transcript, it would be very, very, very efficient. Um, but it wouldn't be anywhere near the quality that everyone was looking for. Um, so if you think of some of the common if you think of the pros and cons of an AI produced transcript, which I'm assuming at this point, probably the majority of you have all used at some point, whether it's personal or work. Um, super, super fast. Right? You put your audio in and usually within minutes it spits out a transcript and you've got a raw a raw transcript of your file. Um, a lot of times the AI does not format. Probably most of the times AI does not format. Um, your transcripts. Um, there are times where it does not not only does it not attribute speaker properly, but it doesn't pick up the correct person was speaking that they attributed attributed it to. Um, some other common errors are mishears um, misspelled The words AI does not necessarily account for context. Um, so you could, you know, someone could be talking about something in a file and it transcribes what it heard, but it's not in the correct context. So therefore it's not the correct word. Um, typically AI does not account for um, paragraph breaks or uh, cadence speech cadence. So those are all things that our authenticators have to think about when they're going through these transcripts. Um, authentication is a word for word. Listen through of the entire audio file. They're not producing the transcript from scratch. Um, but it is a more kind of forensic approach to making sure that the transcript is as close to 100% accurate as possible. Um, so this is kind of just a breakdown of what the project looked like from our end. Um, as far as managing it, um. Our authenticators spent roughly 63% of the entire time, you know, authenticating each of those thousand files or 1500 hours.

19:00 Hanassa Wicks

Um, and this is based. These averages are based on the fact that each file is roughly an hour and a half long. Um, we also had a quality control person put in place. So and I'll talk about this a little bit later. Once the transcripts were received back from our authenticators, they would then go to quality control and keep kind of funneling through the process. Um, so that person spent just over 20% of the entire project time doing that. And then project management, which accounts for, um, you know, setting up the files, moving them around where they need to go. Um, communication between ATC, TheirStory, um, 92nd Street y um, onboarding and off boarding Authenticators recruiting for this project. There was really a lot involved. Um, and again I'll talk about that more. But that's the project management side. And surprisingly, that only accounted for just under 15%. Um, one thing that we did run into, not often out of, uh, 1500 hours or a thousand files, only 20 of them are what we would consider problem audio. Um, that accounts for any audio that is too difficult to hear. Um, audio that might have been in a non-English language, um, which was most of them, I would say most of the 20 files. So Zack is going to just play an example and a quick kind of side story while he's getting this put up. This was the first file when we got this project. Um, when we were getting ready to go, that Michael and I sat down and listened to. And so we pushed play on this file and we thought, oh my goodness, what did we get ourselves into? So I'll let you take this in.

21:04 SPEAKER_S4

Um, and if you can just give a thumbs up when I hit play that you can hear the audio. Uh, and this will just be a quick example of some of those problem audio files.

21:15 SPEAKER_S6

Doctor Kolodny, who's your vasomotor liebfrauen? You mean the force itself and the force of Rome? The room as you. Can imagine. Unlike most. Within a leader, if you were to fail, is a fail. As a. Leader from within. You have the leader in. The. Room. As a leader. As the. Should. Be investigated.

22:41 Michael Sesling

And it's pretty remarkable when you look at the Yiddish that AI was able to transcribe in that. But, uh, and I looked at that transcript initially and said, oh my goodness, what is happening here? Um, but then we realized very quickly, as I thought back to conversations with my zaidy growing up, my grandfather, um, and we were able to quickly come up with a new process to take recordings that were not in English. And I'm going to hand this back to you and as a yes.

23:19 Hanassa Wicks

So and you can see if you were kind of following along, um, in the transcript when the audio was playing, it attempted it did attempt to transcribe it. Kind of. It was almost it's a transliteration of non-English. And then still through some English words in there, I don't know how it decided what it was going to transcribe or not, but yeah. So yes, those were that's an example of the 20, um, uh, problem audio files that we ran into. And in this instance, when we came across these, um, our authenticators just emailed us, you know, gave the file back, Caitlin was notified, and those were replaced by a different 20, um, just so we could keep things going.

24:14 Michael Sesling

So the audio Transcription Center's approach was, um, firstly, uh, just because someone is an excellent transcriptionist does not necessarily make them an excellent authentication specialist. So we needed to come up with a process to be able to test people. Did they have the skill set to be authenticators? And once we came up with, um, with that, we found a recording, uh, that we used and took a AI produced transcript and had folks, uh, authenticate that, read along, listen, and make any corrections. And while, you know, I say that and it sounds like it really should be an easy job, it's amazing the number of people that we had to say, you know, unfortunately, sorry, you do not have the skills it takes to be an authenticator. Um, once we got through that process, I forget the number. Vanessa. Um, I don't remember if we have it in a slide that follows how many.

25:25 Hanassa Wicks

Folks we do have it coming up.

25:28 Michael Sesling

Okay, we'll get back to that in a moment. Zack. Um. We ended up on boarding. Um, I think it was around. Was it 20 folks? Vanessa. Um, as.

25:42 Hanassa Wicks

Yeah. So we we actually only boarded eight new authenticators, like eight brand new authentication specialists. The other 12 to 15 were, um, previous ten, 99 contractors that we had onboarded for previous work.

25:58 Michael Sesling

So we created this dedicated team, uh, going through a whole training session with folks, um, giving them the skills and working with them, uh, to become authentication specialists. And their sole project that they worked on was this 92nd Street Y project that started in August of 24. Was it? And ran until the end of May. Beginning of June of 25. And go for the next slide, Zack.

26:39 Hanassa Wicks

Um. Yeah. So, like Michael had said earlier, we onboarded eight new authentication specialists. The others came from kind of a pool of contractors that we had already onboarded. Um, the test, the authentication test that we sent was a six minute file. Um, we gave them an AI produced transcript that they had to then authenticate and send back, um, one of the one of the things that we really were looking for, because this is not transcription per se, um, and it's more of a forensic approach. So, for example, one of the things that we looked for in the authentication test, specifically, um, was non-speech sounds. The authentication of the actual transcript was fairly easy ish. Um, and then I would look for, okay, well, who added non-speech sounds or who added these certain things into the transcript to make it more readable or more, um, enjoyable for the person who is going through it? Um, one of the at one point in the test file, um, it was a Marianne Moore file, and she's talking, and you can tell that in the middle of her talking, she grabs a cigarette and you can hear the lighter flick, and you can hear her start her cigarette or start smoking her cigarette, and then she keeps talking. So I looked for did who? Who in this process added the non-speech sound of, Um, you know, smokes a cigarette or flicks the lighter. Um, little things like that were really important when going out when going through these test, uh, test files. Um, so once we did that, whoever was onboarded, we did a specific zoom zoom training to go through the different platforms that we use. Um, which I'll get to in a second as well. Um, and just make sure everybody was on the same page. Um. I think that's it for the testing process. Um, again, less than 3% of our authentication specialists for this project were brand new folks. So we had a very dedicated small team of people who worked only on this for a good year. Um, So when thinking about the project management side of things, this is just an example of kind of how things flowed through. We would drop the file into TheirStory. Actually, let me back up. We would grab the file from Dropbox. Um, where Caitlin had uploaded all thousand files, and she broke them down into different folders. So it wasn't just downloading a thousand files at once. We did them in batches. Um, so we would upload a batch into TheirStory. Our authenticators had specific access to their files only within TheirStory. Um,

30:00 Hanassa Wicks

that's where they would authenticate their files. Um. I would it, within TheirStory, move files into folders for them. Once they were completed, I would move them into a quality control folder. Once those were completed, we would move them into a different folder. So it was a lot of migrating of files. Um, which is where we had umpteen spreadsheets. We had a master list. And this list and this list, there was a lot of cross checking, and it was really, really, really important. It seems like it was really tedious or it would be very tedious, but it was super important to do that with so many hours and so many files so that nothing got lost, um, or fell through the cracks. Um, so asana was what we used, um, on the ATC side just to manage our contractors and then QuickBooks for invoicing.

30:58 SPEAKER_S4

And one, one thing I'll add here and then turn it over to Caitlin, uh, to talk about, um, more on the 92nd Street y side. Is that even just managing, uh, files from going from Box.com, where it was originally, uh, into TheirStory? Um, that all takes time and want to make sure that you have the files named correctly and all this sort of stuff that just otherwise would create a lot of logistical overhead and chances for things to be misplaced, lost. And so, although it seems small, one of the first things that we did was build a custom integration with Box.com so that from within TheirStory, you could authenticate using the Box.com login, and then just ingest recordings in batches and be pulling the metadata from Box.com to to TheirStory. So even those small details to try and help as much as possible with uh, the management of uh, even files, some of the, the less, um, uh, you know, fun, uh, aspects to think about, uh, you know, all the dynamic search and discovery that, that will show, uh, is, uh, flashy, uh, and practical and valuable. But even on the back end, uh, thinking about how to streamline those processes, I think was an important part to make this feasible.

32:25 Caitlin Biggers

Yeah. And just just following up because I actually wasn't the person who had to manage these files. We had an internal project manager named Brianna Blackwell who we could not have done this without. Um, but I need to echo again that this whole team making it so easy for us to take the the platform that we were already using and then basically just batch them into folders and not have to do anything special to manage. This was huge. I think anyone who's done an archival project knows it's those little things, um, that make a huge difference in in helping a project succeed. Um, but more to the point, how did the whole project go? Um, so we successfully transcribed and authenticated 1000 recordings, um, over 850 of which are we then published on our website? Um, we plan to add more, uh, if we're able to work out some copyright issues and different things like that. Um, so it can actually grow, uh, if you hit the slide, please. Zack, um, to me, I think the most exciting part of, um, this project really centered on some of the new tools that we employed to kind of further increase accessibility. And Zack gave a little bit of an overview of what we worked on here. Um, but I was really excited about the indexing and the named entity recognition. So beyond transcription and authentication, um, Zack mentioned we also produced an index that breaks down sections of the transcript, um, by topic. So you can say jump to a relevant section, like maybe you only want to hear Arthur Miller read death of a salesman, and you don't care where he grew up, you know, other, other things like that. So you can you can see in the transcript, um, what exactly the speaker is talking about. So you can go right there, both in the transcript itself, and the audio move seamlessly. Um, and I think in the past we had things like generic sections might be introduction remarks. Um, this really lets you kind of scan the index and, you know, generally what the person is talking about without having to listen to any audio if you don't want to. Or maybe you're in a quiet place and you can't listen to anything and you just want to read it. Um, so you can see an example of that. Uh, at the right in our TheirStory player, um, on the 92nd Street Y website. Um, and in contrast to the sample that I played earlier, here's a clip of Allen Ginsberg from 1991 that was not a problem. Just so we could hear, uh, something that worked out nicely. And Zack's going to play that for us.

35:09 SPEAKER_S4

Give me one second as I bring the screen back up. Here we go. Uh, and so here we can see, um, just like Caitlin was saying, this is in the chapter around Ginsburg's, uh, uh, reading of, uh, just Story Road.

35:26 SPEAKER_S6

And, uh, maybe Allen, you recite. Yes.

35:30 SPEAKER_S7

I can remember.

35:31 SPEAKER_S6

Yeah.

35:32 SPEAKER_S7

Millions of children walking, watching the skies, bellies swollen with big round eyes on Jessore Road. I've forgotten. I know it by music. Millions of children watching the skies. Belly swollen with big round eyes on Jessore Road.

35:58 Caitlin Biggers

So yeah, you can see, um.

36:00 Caitlin Biggers

I love clips like this because it's not just poetry reading. It's like someone actively trying to remember on stage and like singing a little bit. And it's. I love these clips because it's a good reminder that, um, you know, this is a record of an actual event that took place. It's not just a reproduction of the poem that you could get anywhere. So that's a little aside about what I find exciting about this collection. Uh, if you can advance to the next slide, that'd be great. Zack. Um, we also built, uh, a named entity recognition, um, or ner pipeline, as Zack mentioned. Um, for those who aren't familiar, as I certainly wasn't when I started this project, NER is a large language model based technology that enables text to facilitate data analysis. Um, basically, the technology uses AI to identify people, places and things, um, and associate them with a standardized concept. For example, our NER pipeline was trained to identify concepts like emotion or nature or even more specific things like city names. And we can then use the labels to identify other recordings in our collection that also, uh, discuss these topics so you can get a sense of themes across the entire collection. That way, even if the person hadn't said that exact word. Um, it can also be used to identify specific points in a given recording. When a speaker touches on a kind of abstract concept like nature, even if they don't actually say the word nature. You might get things when someone's talking about like weather or trees or something like that. Um, so it's really a lot more robust than a keyword search. Um, this slide here shows the portal. And at right you can see the concepts, um, labeled and highlighted in different colors. This is static. Um, and I have multiple labels active right now. But if you look at the bar towards the bottom, if I was to select something like only emotion, uh, as the, as the concept that I wanted to view, the bar at the bottom would show colored dots at all the points along the specific recording. Um, when an author touches on emotion and you can do that for any of the terms that, uh, it identified. So to me, it really helps parse an audio file and kind of understand the whole thing in terms of, um, concepts, without having to listen to the entire thing and make notes on what authors talk about yourself. Um, so that's huge. Uh, between the authenticated transcript, the index and the Na data portal, we aim to kind of help people get into these recordings. Um, and I think that has been pretty successful. Um, we've, we've since this has been online for, uh, you know, more than six months now, we've had some press and then we've had really positive feedback. People love that. You can use these chapters in the different ways that you can research the collection. Um, and I've heard that it helps people kind of make connections across the collection as a whole. That would be very difficult to with, with just single videos up online, something like that. Um, so we hope that we've produced something that preserves these incredible literary voices and just generally helps people use them. Whether you're a professional researcher or someone, um, who has a casual interest in in poetry or the why or any other topic that this might touch on. Um, so I'm not the expert on GnRH. So if Zack wants to add anything on the research portal, I think you should certainly do that here. Um, but yeah, that's that's my spiel about our project.

39:36 SPEAKER_S4

Yeah, absolutely. I think the only thing I would add, you know, I think you you nailed it is, is, uh, you know, I can just show it in action. Um, uh, very, uh, very quickly, um, to kind of bring to life a little bit of what you just summarized. So, um, let me pull back up what we had before. Okay. So, um, here is the the research portal. You can access it if you go to the 92nd streetwise um, uh, main page. And if you go to the research portal here, uh, you can also just directly from their website, access the, the recordings. Um, but if we go to the research portal, maybe I want to find any place where, uh, where a particular person is mentioned by name. And maybe I'm interested in something like, you know, what was the role of women in literature? And normally we're kind of beholden to the specific words that are spoken. If we're doing search or it has to be in the the metadata that has been tagged, uh, manually. But if I don't really know what's in here, how might I just be able to use my own natural language? And so, um, if I just search here across the 850 plus recordings just that fast, I can find these thematically, uh, relevant, uh, excerpts. So here in a Q&A between the audience and the speaker. One of the audience members says, you know, one of the themes that emerges in a lot of the criticisms of your writing, just observations of your work, is the role of the male in a lot of it. And of course, you've argued in The Color Purple, and it wasn't really about men, it was about women. And then it goes on to mention this particular person. And so I can now click on this excerpt and go directly to that exact moment in the recording of the. And I can also then explore by those entities that, um, that Caitlyn was talking about. And so we can see that full list that's been extracted, um, in this recording. But maybe I want to find all the places where Toni Morrison is referenced, whether that is in this individual recording here, or maybe it's across the collection. And this is, this is what, uh, Caitlyn was talking about, you know, helping people navigate the collection of content, not just a single recording, but especially with research that's about what's similar, what's different, compare contrast. Uh, I maybe want to be able to, to see what the individual recordings in total tell us that a single recording wouldn't be able to alone. And so maybe I can go in here and find where Toni Morrison is referenced in other places, and you can do this across any of those types of entities. Or as we were showing just a moment ago, that just kind of natural language search. Um, now, one thing that I'll add is what we have have done, um, one is release a number of these tools as open source. So both the, uh, machine learning model for extracting named entities, we've, we've released, um, this has been, uh, starred at the very least, if not more so adopted by folks at, uh, Adobe, Microsoft, uh, AWS, um, uh, Yale, Duke University are all people that, um, have, uh, at the very least, uh, been paying attention to this and are starring this, uh, this platform, a number of people who are forking it and using it on their own. And we want to make these tools available to anyone, not just if you're using the TheirStory platform but but generally and the same thing. Um, this um, kind of interface for uh, being able to make your audiovisual collection searchable. Uh, we have also um, released as, uh, as open source. And so this was done just this past week. This would not have been possible without, uh, the commitment from the 92nd Street Y from melon. Um, and, uh, and everything that Michael and his team did to make this project successful, um, and, uh, and with kind of the learnings we had. And so we now have this available, I'll put the link in the chat, um, as well, where this package is up, uh, kind of the best of, uh, industry leading open source AI native databases that allow for that kind of thematic search. It also allows for things like having ChatGPT like conversations with collections. It was interesting. We had implemented that initially, but decided not to release that as a part of this portal just because we didn't want, um, AI interpreting, uh, works of creative, um, uh, individuals. Um, but all of that is essentially packaged up and, and something that now others can use, uh, really, really simply. Um, and so if anyone's interested in creating this type of kind of custom branded, uh, research portal for, for themselves, then, um, we'd be happy to, um, to have follow up conversations. I'll put a survey, um, into the chat as well. So if you want to have a follow up conversation with Caitlin, with Michael, with Vanessa, with myself on any of this, uh, we're happy to to chat. But this is all in terms of what you can build because it's an open source template. We have folks that have customized it in a number of different ways. Um, and so it's it's really, um, amazing. I'm excited about what's, what's possible. But again, this would not have been possible to get to this point without the 92nd Street Y and the the trust to. To build something like this together. Um, so, uh, I'll pause there, just on the technical sides, let me share my screen back to our, uh, to our slides and to some of our closing takeaways.

45:26 Hanassa Wicks

To me first, isn't it? Um. One of the big takeaways from this, obviously, authentication is key when using transcripts that are produced by AI. Um, you figure out what your goal is, figure out you know, what you're looking for as your end result. And then from there, you can decide what kind of authentication you need or how in-depth you need. Um, but I will say after this project and others. Um, I have yet to find an AI produced transcript that is to my 99.9% liking.

46:19 SPEAKER_S4

And, uh, another thing that I think is important takeaway is, um, understanding when to use AI, when not to use AI. Um, and this comes down to both things, like organizational policies. It comes down to ethics. It also comes down to sometimes AI in and of itself is not the solution. But, you know, an integration with Box.com, something very simple that's more about just streamlining the process in this case. But I think it was a key thing where as someone who's building the technology to really listen to the groups that we're working with, you know, I was personally all excited about the, you know, ChatGPT like conversations that you could have with the collection and and not that I think, uh, you know, Caitlin wasn't, but but we had to listen to the idea of, like, there could be legal concerns, um, with making that publicly available. And, um, and even if it could aid significantly in parsing, you know, 850 plus recordings and finding out what's in here and how to navigate to those moments really easily. Um, it's interpreting, you know, a creative person's work. And at the very least, that might not be accurate. Uh, or it could, you know, Caitlin would have to go back and talk to 850 plus authors, some of which you might not even be alive to understand, if that's something they'd be okay with. So there we have to be cognizant of the ethics, uh, around all of this and listen to the people that were serving and, um, and make sure that what we're building needs meets their needs. I know, Caitlin, if there's anything that you would want to add or modify to, to what I, um.

47:56 Caitlin Biggers

Said there no one, just 100% agree. Yeah, it was really, really cool. And I wish you know that in a perfect world, we could make that public. But I do think it warrants further consideration. Um, so yeah, and Zack was willing to work with us on that, of course.

48:11 SPEAKER_S4

Absolutely.

48:14 Michael Sesling

It was amazing to be able to partner with TheirStory. Um, who was the technology and the firm that I worked for that had the humans who were able to get all this work done and find that balance, um, where the technology and the humans could meet and make this, um, product, shall I say, uh, for the 92nd Street y that's proven to be so useful and has now been viewed quite a bit. Uh, Caitlin, if I'm correct. Um, so it was just amazing to be able to find that connection between humans and technology and find a way to continue to improve, and there are projects that still have to be all human for a number of things. As as Zack was mentioning, um, earlier. Um, but it's important to find those places where there's that intersection, um, where we can improve that process. Caitlin.

49:16 Caitlin Biggers

Yeah. And just finally, as the screen says, trust. Trust is everything. Um, I think anyone who has worked on a large project, but more specifically a grant project, you, your relationship with your vendors is absolutely everything. Um, you need to be able to trust that the work is going to come in quality and on time. Um, when working with NASA, I would ask a random question because I knew she was more on top of the spreadsheet. And even if there was, you know, as there inevitably is when you're dealing with somewhat large, um, pieces of data, you know, if there's a discrepancy, I'd send it to NASA, and 20 minutes later I'd have an answer about what went wrong. And we found this this issue immediately. And I think, um, I'm sure many of you have worked with other vendors where you're you don't get that kind of service. And sometimes they're like, I'm not sure we'll see what happens. Um, so I yeah, I have to just stress that trusting the people that you work with is really, really critical and, um, trusting that they have interesting ideas and we should pursue these, um, kind of things, like, like what Zack was talking about. Um, the research portal was not even considered as part of the initial scope of the grant, but we had to trust this is going to this is a natural extension of, of, um, the type of work that we're trying to do here. So, yeah, I really just think, um, a takeaway for anyone embarking on a project is make sure that you vibe with the people that you're going to work with and, and can trust them 100%.

50:49 SPEAKER_S4

Uh, so we have a QR code here for if you want to access the the collection through the 92nd Street Y site. Definitely encourage anyone to take a look at that. It is just such a unique and interesting and fun collection to to listen to. Um, and uh, I have also dropped into the, uh, chat. Um, one is a Google form, so if you want to follow up with any of us, um, uh, individually, uh, whether to learn from Caitlin, uh, you know, as an archivist, what was this process like? Um, Michael and Vanessa around? Anything on the human capital side? Uh, or, uh, with me about TheirStory or making use of any of the open source materials? I've also put links to the open source, uh, research portal, uh, and also to the open source um named entity recognition model as well. So thank you everyone. Happy to answer any other questions that folks might have we can stick around for for a little bit. Um, and really appreciate everyone's time today.

52:02 Michael Sesling

Thank you all.

52:03 Hanassa Wicks

Thank you.

52:04 Caitlin Biggers

Thanks. Looks like there were a couple of questions, but I think you might have answered them. Zack, in the course of, um, the presentation, but I think.

52:31 SPEAKER_S4

Yes, I see a few things here. Uh, so one, uh, Andriana asked about using the named entity recognition model within TheirStory. Um, we have a mechanism for being able to do that. We should have a one on one conversation. So, Andriana, I'll send you a note, uh, afterwards, and then we can talk about getting that, um, what that can look like. Uh, let me see. Um. Uh, Benjamin. Uh, Rubin. Uh. Hi, Ben. Um, asked about, uh, budget. Uh, let's have a conversation directly. That depends on scale, um, of the volume of recordings that you may have. Uh, and. Yes. And I will, uh, we will share the recording and and slides. Well, I think from here, unless there are any other questions, we can call it a wrap. Thank you. Everybody

53:46 Caitlin Biggers

Everybody good? Yeah. Thank you.

53:49 Michael Sesling

Awesome. Thank you Zack.

53:50 SPEAKER_S8

See you guys.

0:00

About TheirStory

Ask about the interviews

About TheirStory

Ask about the interviews

Ask about the interviews

Ask about the interviews

Project Background and Initial Challenges2:00Michael and Zack discuss the initial meeting and collaboration for the transcription project of the 92nd Street Y's archives, including the challenges and importance of accurate transcription.

92nd Street Y Archive Overview4:36Caitlin provides background on the 92nd Street Y, its historical significance, the archive's content, and the importance of making it accessible to the public.

Project Details and Grant Process8:47Discussion about the project's scope, the transcription and authentication of recordings, and the grant proposal process with the Mellon Foundation.

Technology Integration and Solution Strategy12:00Zack and Michael elaborate on integrating TheirStory's platform with existing systems, using AI for transcription, and the technical improvements that made the project feasible.

Process of Authentication and Human Involvement15:20Hanassa discusses the process of transcript authentication, the role of authenticators, and the challenges faced in ensuring transcription accuracy.

Managing Problematic Audio Files19:00The team discusses how they dealt with problematic audio files that were difficult to transcribe, including a specific example of a challenging Yiddish recording.

Project Workflow and Management Tools24:14Details on the workflow and management tools used to manage the transcription project, including Dropbox, TheirStory, and Asana.

Research Portal and Access Innovations30:00Caitlin and Zack introduce the research portal developed to enhance access to the 92nd Street Y's collections, featuring indexing and named entity recognition.

Open Source Tools and Future Applications36:00Discussion on the open source tools developed during the project, and their applications for other institutions, as well as the ethical considerations of using AI in archives.

Key Takeaways and Lessons Learned45:26Closing remarks on the importance of collaboration, trust, and balancing technology with human input in archival projects.