WEBVTT 00:00.000 --> 00:13.440 So this is our last AVG Devroom talk of the day, it's another podcasting one, I'm using AI to do 00:13.440 --> 00:18.600 that by guessing it's, well, I find out how you can use AI, thank you very much. 00:18.600 --> 00:23.640 Hello everyone, today I'm going to talk to you about a new project which is called Pod 00:23.640 --> 00:36.640 Libre and it's a podcast audio editing software for the AI age. 00:36.640 --> 00:43.280 This is the mascot for Pod Libre and we have some stickers so feel free to come and get 00:43.280 --> 00:45.280 some after the presentation. 00:45.280 --> 00:52.440 So let me introduce myself, I'm Benjamin, I love podcast, I love free an open source software, 00:52.440 --> 01:01.560 I work at Lina Goha, developing truly open source AI solution and I emphasize on truly open 01:01.560 --> 01:02.560 source. 01:02.560 --> 01:07.440 I'm also the founder of Adorece which is a company that develops open source solution 01:07.440 --> 01:15.040 for podcasters, one of them is CasterPod which is an open source podcas hosting platform 01:15.040 --> 01:23.200 and I'm also a photographer for Wikipedia which does nice picture for Wikipedia. 01:23.200 --> 01:33.040 So first let's talk about who is a developer in the room, okay that's over 100% and 01:33.040 --> 01:44.000 who is a podcaster, okay like 25, who wants to be a podcaster, okay and who is listening 01:44.000 --> 01:49.760 to podcasters or like 100% okay. 01:49.760 --> 01:55.760 So now I'm talking to the person who are already podcasting what software do you use, what 01:55.760 --> 02:01.280 is your workflow, like the main editing software. 02:01.980 --> 02:05.580 Oper, vapor audacity. 02:05.580 --> 02:11.160 Logic okay, that's it. 02:11.160 --> 02:24.000 No one is using outer, okay, like the rhythm, okay so let's start with a definition 02:24.000 --> 02:32.200 what is put lib or per lib array of its it's come from podcasts and from lib which 02:32.200 --> 02:40.920 is free as in free speech. So it's either French or Spanish. I'm taking just two 02:40.920 --> 02:47.560 examples that I wish we'll see in the dictionary in a few years. I switch to 02:47.560 --> 02:52.320 per lib because I want it full control over my podcast workflow without sending my 02:52.320 --> 02:57.480 audio to the cloud or the per lib approach means owning your entire 02:57.480 --> 03:05.120 projection chain from recording to RSS feed. Because of course we all agree 03:05.120 --> 03:12.080 here that podcasting is linked to RSS. If there is no RSS feed then it's not a 03:12.080 --> 03:20.280 podcast. If it's on YouTube it's probably not a podcast. So I'm going to tell 03:20.280 --> 03:28.920 you my story. Bonnie just before me said that everything about recording a 03:28.920 --> 03:33.440 podcasting is like a religion. And I totally agree with her. In fact I 03:33.440 --> 03:42.600 couldn't agree more. So welcome to my church. Back six years ago we wanted a 03:42.600 --> 03:47.920 solution to host our podcast and to host the podcast for our customers, our 03:47.920 --> 03:53.080 clients. And we couldn't find any and more of her we wanted one which was 03:53.080 --> 03:59.200 connected to the Fediverse and that was we could find even less. So we started 03:59.200 --> 04:05.520 CastoPod as a side project. We're deeply grateful for the community that 04:05.520 --> 04:11.960 we've built like over 1.7 million hours of content. So people are trusting us 04:11.960 --> 04:17.360 with their content. Most of them are self hosting which is what we wanted in 04:17.360 --> 04:23.080 the first place. So that's really great. And it's translated in over 30 languages. 04:23.080 --> 04:30.480 Some of them we are like the only podcast platform having these languages like 04:30.480 --> 04:39.680 Roton. In France there's no other platform using that language and this is why 04:39.680 --> 04:45.240 we love open source because if you want something you can have it. So we're 04:45.240 --> 04:57.120 our podcasters too. So this is my church. Over the years I started using logic 04:57.120 --> 05:03.200 then other because it's open source and because it's non-destructive then I 05:03.200 --> 05:09.080 switch to audacity because it's easier faster and it's almost non-destructive 05:09.080 --> 05:13.800 now not 100% but almost and with the new version you can use it with pipe 05:13.840 --> 05:22.200 wires. So yeah it makes sense but we are a big advocate for transcriptions, 05:22.200 --> 05:31.920 for obvious accessibility and discoverability issues. So we need to make transcriptions 05:31.920 --> 05:37.720 and I developed over the years some patent scripts to do that and many scripts. 05:37.720 --> 05:46.560 So publishing an episode has become very cumbersome and when talking to other 05:46.560 --> 05:51.800 podcasters we will realize that we are not the only ones with these issues. So this 05:51.800 --> 05:59.320 is why we decided to start public labor. One other thing is that the software we are 05:59.400 --> 06:08.120 using so we per our audacity to read them logic. They are done for musicians and 06:08.120 --> 06:12.640 there are many many things that we don't need and many things that we need that's 06:12.640 --> 06:22.200 aren't there. So the workflow is not very appropriate for podcasting. So this is an 06:22.280 --> 06:32.760 example. So this is my podcast the latest episode yes. So as I said and as Bonnie said 06:32.760 --> 06:38.680 this is my religion maybe yours is totally different and this is why I'm talking to you 06:38.680 --> 06:44.520 now because at some point we need some feedback because there is no way that we have 06:44.520 --> 06:50.440 all the good answers just with a bunch of us. So we really need your feedback if there is 06:50.520 --> 06:57.640 one thing that should come out this stock is feedback from you. So here you can see there are 06:57.640 --> 07:09.160 some speakers so one microphone per speaker you can see that when someone is not talking it's totally 07:09.160 --> 07:19.400 blank it's muted. Here there are chapters you can see yeah next weekend it's first them here 07:20.440 --> 07:30.200 the Trump is the best advocate for open source software here is the music and doing that 07:30.200 --> 07:36.920 that's that's pretty simple it's very easy they are like four tracks plus the chapters 07:39.400 --> 07:46.200 that should take 10 minutes and I really do think that's doing this editing this should 07:47.080 --> 07:58.360 no more than 10 minutes but the truth is it takes hours because you need to normalize the level 07:58.360 --> 08:06.760 and to compress the audio and maybe here this was our intern he's 14 so he spoke too far away 08:06.760 --> 08:16.040 from the microphone so the ratio signal noise was not good enough so we have to level up the sound 08:16.200 --> 08:23.080 so there was too much noise so we had to remove noise but all these should be almost automatic 08:23.080 --> 08:31.080 it should be most complicated part you have to do which should be checking a checkbox because 08:31.080 --> 08:40.840 these are the thing that we do every week this is the transcription edition if you do we do 08:40.920 --> 08:47.240 automatic the transcription but you have to correct it and there is no software that allows you to 08:47.240 --> 08:57.720 correct that in a way that you feel your in 2026 and more of the transcription is done afterwards 08:58.840 --> 09:03.400 but I needed before editing because the transcription could be very useful for editing 09:04.280 --> 09:11.960 so this is a benchmark that we did over a year ago with many many tools open source and non-open source 09:13.320 --> 09:23.720 I'll publish the the full presentation afterwards and none of them answers our needs 09:24.600 --> 09:32.600 our needs so we need transcript driven editing because we want to be able to use the 09:32.600 --> 09:39.080 transcription to edit and to know which one should be taken out noise reduction one metadata 09:39.080 --> 09:46.200 place for everything because there are many metadata the title the show notes the transcription 09:46.200 --> 09:52.040 the chapters and right now you have to put them like maybe there are some ID three tags within the 09:52.120 --> 10:01.160 MP3 file and in the end it takes much time for no ID value chapter management etc 10:03.160 --> 10:12.280 so we asked you is there anyone here who answered this poll in the room no you don't remember so we asked 10:12.280 --> 10:18.920 that's on the the fetivers on mastodone and we asked would it make sense so it was over a year ago 10:19.000 --> 10:29.480 we did make sense to develop a tool for podcast edition and yeah we we got like 45 plus 36 10:29.480 --> 10:40.440 percent we said yeah so put libre is a podcast editor built by podcaster for podcasters and open source 10:40.600 --> 10:49.400 so the key principles we want something that's cross platform because even though I'm using 10:49.400 --> 10:58.120 linux and I think everyone should one step at a time it should also be able to run on macOS 10:58.120 --> 11:05.320 and windows privacy focus because we don't want to sense our data to any big tech we need a 11:05.320 --> 11:13.160 plugin based architecture because I know that my workflow is not yours and even my workflow 11:13.160 --> 11:17.880 can evolve and change depending on which podcast if I'm hosting two podcasts they probably 11:17.880 --> 11:25.560 don't want to have the same workflow of course it's open source and to be able to run offline 11:25.560 --> 11:33.560 I don't want an application that's used APIs for everything maybe there's one plugin that we 11:33.560 --> 11:42.680 have an API and run on the cloud but I want it to be able to run offline so this is the like the 11:42.680 --> 11:51.080 basic workflow that we designed so you create a podcast then an episode your import your files 11:51.080 --> 11:59.240 maybe later but that's not our priority your record from the application and then you would 11:59.400 --> 12:06.760 like normalize audio reduce noise compress then transcripts and you'd be able to correct the 12:06.760 --> 12:15.080 transcription and go back and then serve the music something that's really miss on all 12:17.400 --> 12:25.720 audio editing software is the ability to have a music which length can adapt to the lens of my 12:25.720 --> 12:33.000 introduction so like we've beginning a loop and ending and I would just want to be able to 12:33.000 --> 12:39.000 drag and drop and say use the loop as many times as possible so that's the intro music last 12:39.000 --> 12:46.600 as long as my introduction my spoken introduction then you did the chapters the metadata and 12:47.400 --> 12:58.200 you export you publish to whichever hosting system you want to so this is something that 12:59.480 --> 13:06.600 it would look like this is a mockup it doesn't exist yet we have some stuff running but I want 13:06.600 --> 13:15.240 to be showing them to you today because we a little bit behind schedule and this is what we 13:15.320 --> 13:22.760 started working on so we are using Python because I said before I developed many scripts with Python 13:22.760 --> 13:29.320 because we need transcription we need many things that are working there where you can p-y cute 13:30.280 --> 13:39.080 so that it can run on any platform we spur a set it and noise redirect by dub and plugin architecture 13:40.040 --> 13:47.160 so we've a little more detail this is what it looks like I published that so that you can have a 13:47.160 --> 13:55.400 look at it and give us feedback of course because as I said and I'm repeating myself we need your feedback 13:57.480 --> 14:06.040 so we built the very first break we have a work in progress we have plugins we have transcription 14:06.920 --> 14:15.960 we started editing the audio so we have the core architecture the plugin system is functional 14:15.960 --> 14:21.880 we also got a grant from an alnets to make sure that this is sustainable and that's will we 14:21.880 --> 14:33.480 go to the end my company address of course is standing it but what this means it's things will break 14:34.440 --> 14:43.960 we will the UI will change your feedback shapes everything repeating again and now it's probably the 14:43.960 --> 14:50.440 best time to get involved because the more the mayor year but also the better ideas we can get 14:53.560 --> 14:58.520 we know that it won't be that easy there are many many things to do 14:59.400 --> 15:08.200 the platform compatibility GPU acceleration we don't even know if we want to do something like that on 15:08.200 --> 15:18.680 macOS with the new silicon systems managing compatibility with Python because plugins is the key 15:18.680 --> 15:28.120 and you know that it's very easy to make a mess with plugins and with stability because I'm using 15:29.160 --> 15:39.560 audacity I'm using like the latest version 3.7 and it's not stable and I say a lot of bad 15:39.560 --> 15:47.800 words when I'm editing right now I haven't tried the version 4 which is not finished yet you can 15:47.880 --> 15:55.800 do chapter is it's missing a lot of features but we need something that's stable and we want to 15:55.800 --> 16:00.840 spend like 10 minutes maybe half an hour on it not hours because it keeps on crashing 16:04.440 --> 16:12.840 so if you want to join us if you want to give us your feedback it can be technical it can be 16:12.840 --> 16:19.640 we're talking of features and environments things that we haven't sort of there's a website which 16:19.640 --> 16:28.200 is put lib.org there's a community website which is community.podlib.org and the source code is on 16:28.200 --> 16:34.760 codeberg because we don't like GitHub because we don't like Microsoft. 16:38.600 --> 16:41.480 Thank you. 16:51.800 --> 16:57.800 Thank you very much questions if we can write there's one right up at the back we have to use 16:57.800 --> 17:14.360 the microphones that people online can hear so thank you for everything I'm an enthusiast 17:14.360 --> 17:23.240 customer the user in self-hosting and my question is exactly if in your roadmap there is any 17:23.240 --> 17:31.480 seamless integration to start from the recording to the publishing on one very best platform which 17:31.480 --> 17:43.800 may be cast upon the world. Yeah of course we in the very first version we plan to integrate 17:43.800 --> 17:50.600 with cast-op-odd so that's you can publish directly from pod lib to cast-op-odd without even opening 17:50.600 --> 18:01.480 Firefox we also plan to have a local publishing because I think that a podcast in the end it's just 18:01.480 --> 18:10.680 MP3 files and analysis text file so we should be able to publish to a static website so this is one 18:10.680 --> 18:16.760 other if you have a static website also we want it to we want to the the app to be able to 18:16.760 --> 18:22.280 upload it to an FTP SFTP server so that's win one click it's online and you don't even need 18:22.280 --> 18:31.000 PHP on your server one thing that we also integrate and there's a talk later this afternoon 18:31.000 --> 18:41.480 it's called Fur Camp it's also a project granted by an alnet and it's a software that 18:41.480 --> 18:51.960 publishes static website for audio so and the thing is we're doing a plugin architecture so 18:51.960 --> 18:58.280 that's when you create a workflow you select the plugins that you want and in the end if you want 18:58.280 --> 19:03.640 to publish to cast-op-odd you select the cast-op-odd plugin and you say this is where I want to publish 19:04.200 --> 19:10.280 and if you do another podcast on another platform you duplicate this workflow and you just 19:10.840 --> 19:17.880 change the publishing plugin at the end put to publish it to I don't know Spotify for podcasts 19:19.080 --> 19:37.320 or YouTube podcast that's your decision and I have questions about the workflow you showed and about 19:37.320 --> 19:48.200 the software so I'm working in a okay I mean a nonprofit radio in Britain in France and we are 19:48.200 --> 19:56.760 looking for software who are more easier than a rapper so I'm very interested about the 19:56.760 --> 20:04.280 little although I'm also a sound engineer and I've seen a I have questions about some of the 20:04.280 --> 20:12.600 technical aspects and especially your work for you mentioned compressing audio before editing 20:12.600 --> 20:21.400 in and something like that I just want to be sure if you in the software you if it will be as flexible 20:21.400 --> 20:32.120 as a rapper which also a rapper can not only do music but a very lot of stuff I've done mixing 20:33.000 --> 20:41.640 for short films and other stuff so maybe we can also meet later 20:43.640 --> 20:48.680 after I'm not a big fan of reaper because it's not open source and everyone is using it because 20:48.680 --> 20:58.440 it's free if you choose not to pay for it so I don't like this business model either you pay for it 20:58.440 --> 21:03.720 or you don't and many podcasters in France are using it because you can cheat and get it for free 21:03.720 --> 21:14.520 without but and it's not open source and it runs on Linux but with some tricks the thing is 21:14.520 --> 21:22.840 to answer your question the workflow that I showed is an example it's my church and if you don't 21:23.720 --> 21:33.560 a few months ago I didn't even compress sound because I like sound like writer of the microphone 21:33.560 --> 21:39.240 I take very good care of the environment and the microphone so that's I don't have to change anything 21:39.240 --> 21:45.640 and some listeners say you have to compress because in the car we can't hear and so I change my mind 21:45.640 --> 21:53.240 and I change my workflow there is not one workflow even for a single person so of course you have to 21:53.240 --> 22:02.120 be able to adapt so very nice idea very nice initiative I think so thank you for that I have 22:02.120 --> 22:08.120 two questions but I'm not sure if I'm allowed the first one is about the PyTorch dependency that you mentioned 22:08.120 --> 22:14.120 that typically that's used for training not for inference and one because I know from experience 22:14.120 --> 22:19.960 dealing with packaging Python stuff course platform that becomes very hard to manage is that something 22:19.960 --> 22:28.920 that you are going to maybe require at runtime or only for like building or so yeah this is a 22:28.920 --> 22:36.600 copy paste form everything that we've been working on the past few years with my colleagues and I 22:37.400 --> 22:46.840 there's one thing that we not we don't know for sure maybe at some points we will build several 22:46.840 --> 22:53.400 versions so that you can choose maybe you don't care about transcription and you don't want this 22:53.400 --> 23:00.200 at all and you want something like a minimal install very small so that you can run on the very 23:00.280 --> 23:06.680 old laptop and that's probably what we're going to do if you want to keep it simple to provide 23:06.680 --> 23:16.040 several versions and we'll build many of them the thing is it has to we don't want too many of them 23:16.040 --> 23:22.040 because otherwise you won't be able to choose which one you want if I'm allowed the second question 23:22.040 --> 23:28.680 that's good the other question would be regarding because I saw their deep filter net that I had 23:28.680 --> 23:34.120 some experiments done with and there is a way to package everything it's something that I'm working 23:34.120 --> 23:38.120 together with a friend of mine to package everything into an audio plugin so you don't have to 23:38.840 --> 23:43.800 sell up any training don't have to sell up dependencies it just ends up with a single plugin binary 23:43.800 --> 23:49.960 basically would you consider maybe picking out all the pytorch all the other dependencies if you 23:49.960 --> 23:57.000 can just install one single plugin that does for example the like vocal isolation if it does it for you 23:57.720 --> 24:03.240 do you still need all these other AI tools if you have plugins that can do the real time processing 24:03.240 --> 24:11.960 of the audio we want to put as much as possible into the plugins we are I'm currently using whisper 24:12.840 --> 24:23.400 but whisper is not the the only transcription engine in town there is not a skewty the even 24:23.400 --> 24:31.800 meta release the like months or two ago trend models for a 1000 languages we've translation so 24:35.720 --> 24:48.600 we want to put as few as less possible in the core app of course unfortunately we've run out of time 24:48.680 --> 24:53.880 and we keep very strict time at fostered so thank you very much great to all great questions 24:53.880 --> 24:59.560 and great AVT creation everyone as our first