Fish Audio S2.1 Pro Full Test – Building A Video Translation Pipeline!

0:00 Hey guys and welcome back to the channel. Today I put that in the intro. All right, let's just try with both English.

0:11 Today we're going to be testing Fish Audio S2.1. Today we're going to be using Fish Audios S2.1 Pro speech model to clone my voice, turn it into a different language, and then allow me to

0:23 take back control of some of my videos that have been getting uploaded on other platforms. though for a bit of prerequisite and some background information as to why this is necessary.

0:33 Let's just take a look at this clip which in my opinion is a bit of a problem. So this is a website called Billy Billy and it is a Chinese website where videos get posted. You can think

0:42 of it almost somewhat like YouTube. Now, the biggest issue here is that I did not upload this video and there are many other videos of mine uploaded on here from various accounts that have been

0:52 doing said uploading and a lot of other creators who make videos in this niche like AI model reviews, things like that are also on here. Um, so let's just listen to the cloned voice of mine here

1:03 that is cloned into Chinese and just listen to this. So deep. So deep. >> So the actual clone itself, someone's

1:23 taken my voice and translated it into Chinese and then like overlaid it back into the video to upload it here. >> The actual regular speech is not bad at all, but it is like portions like this

1:36 in the introduction. Yeah. So, I don't even want to laugh now because then if it gets translated back using this same person's workflow, it's going to sound like that. So, today's

1:50 whole video is essentially going to be this. We're going to take one of my videos, in specific, the Quen 3.8. Next video, we're going to have a local AI workflow that actually slices it up

2:01 based on timestamps, transcribes the speech, and then using Fish Audios S2.1 Pro and either the API or the MCP server, we're going to have that video be programmatically changed into my

2:14 voice cloned into Chinese. And then that audio will be pulled back down, overlaid into the video, and then I will have my own videos to upload on Billy Billy that I myself have uploaded and created. So

2:26 this is going to be partially a local workflow and partially a cloud workflow just using fish s2.1 pro. Now I have actually tested some of their texttospech models I think a bit over a

2:36 year ago. So this video from over a year ago was actually testing one of the fish audio open-source texttospech models. Almost ironically in the introduction I showcased my clone voice being

2:47 translated into a different language. So perhaps that was just kind of like a precursor to what we're going to be doing today. So, with that, let's just jump in and get started. I generally

2:57 don't have a fully concrete plan in how I want to specifically implement every part of this stack that's going to produce these videos for us. However, I have like concepts of a plan, which is

3:08 enough to get started. So, before we get into it, I do want to say thank you to Fish Audio because they were kind enough to support me in my quest to basically try to take back some of my videos that

3:17 have been getting lifted and posted here with varying levels of voice cloning success. Additionally to that, they were kind enough to basically just allow me to use their platform for this. So,

3:25 thank you to Fish Audio for the support. Additionally to that, at least for the duration of the next 48 hours, there is free access to the S2.1 Pro API for this model for folks who want to test it,

3:37 assuming that it falls under the fair use policy. So, with that, and depending on the time that this video is posted, this number may be different up here. However, this is going to be fun because

3:46 text to speech videos I have not done in quite a while. If you are an OG to the channel, thank you. And you may be familiar that some of those were very, very, very entertaining over the past

3:56 couple of years we've tested them. So, let's get started with step one, which is actually cloning my voice. So, our first step here, just from the fish audio dashboard, I'm going to hop into

4:05 create voice, and I'm going to opt for the instant voice cloning feature. Now, I have pulled a specific clip from the Quen 3.8 next video that I'm actually going to be using as a demonstration of

4:17 this endto-end pipeline. I believe I have about 15 seconds of audio ready to be posted there. Let me take a peek and see if I recall. So, this is the clip that I will be using for the cloned

4:27 voice. So, we're now doing a true local test with this at a Q4KM quantization running on the machine behind us. This is the unsloth quant and we've given it the subway FPS test. I want to give it

4:38 something relatively simple to accomplish just to see how this performs in this quant. So I think that's about 15 seconds and it has enough pertaining to like the type of speech like that

4:50 things like quant and local and those sorts of words although I don't know how much that will actually affect this. And we now see that it is uploaded here. We have 15 seconds uploaded. 30 seconds is

5:01 enough for most voices. Longer audio can still help. So we see that's enough. I wonder if I should upload a bit more. I want to just see how it does with the 15 seconds. but we may go further and have

5:12 a more intense voice clip depending on the quality we get. So, let's see what we get right here just from 15 seconds. And it is currently analyzing the audio. All right, it does say good right there,

5:22 which I would assume is good. So, next step, we're going to just continue to details right here. And this is all right. Good. It automatically named this voice for us clear tech narrator, which

5:34 thank you. But our next step here, we're basically going to be able to change some things depending on what our purpose of using this voice is, which for this video, of course, is probably

5:43 in scope as a clear tech narrator. However, this is also where you can do perhaps a little bit of entertaining stuff with this, which we may have to try. Middle-aged. Okay, I don't know how

5:52 I feel about that. Now, I do also have some options here, such as selecting specific use cases. Narration and educational are fine, but I do also want to include social media. I'm kind of

6:02 tempted to include entertainment as well, just because some of these videos are perhaps leaning more on that side than like educational. And with that, we also have some other options in terms of

6:12 how specifically we want to handle access to this. I'm going to keep it private so it's only accessible to myself or else we may have just like more instances of things like this

6:21 happening that I myself have caused. So with that, our next step is just going to click create. And you have to confirm obviously you have permission to create and share and clone this specific voice

6:31 which I have granted for myself. So we're cool with that. All right. So let's take a look now. Let's just Here's an audio sample. And this is we are performing a true local test with our

6:42 Q4KM quantization running on the machine behind us. We've given it the subway FPS test aiming to see how it performs in the specific quant. Now, I do believe that's actually the clone of the voice

6:53 just playing back what was transcribed in the speech segment of the sample clip. It's actually kind of hard to tell. So, that's a good thing, I guess. Now, if we wanted to make a few changes,

7:02 we can also click on the edit model button right here. And if we scroll down, we see that there are a bunch of selectable voice qualities we could have. These have been auto selected just

7:11 based off of what it assessed based off of this clone and when it analyzed it to see. So, I'm fine with all of these and I don't think I'm going to change anything in terms of that because the

7:20 voice seems fairly good as is. So, the next step, I think, is just going to be to click use voice and we're going to actually be able to play with this just a bit. We can see we're using the free

7:29 version of Fish Audio S2.1 Pro and that will suffice for today's section. So the next thing I would like to do now is to basically again you can search back on the channel

7:42 some of the text to speech videos were hey how are you oh that was the speaker I apologize hey how are you subscribe and we have two different

7:54 samples here just to do like an AB test how are you subscribe >> I think one is better and I'm probably okay with that so I've translated just a

8:04 simple sentence right here into Chinese. And I'm going to copy paste this now into fish audio. This is something that when the pipeline is built to autonomously transcribe the videos,

8:14 translate them into Chinese, and then send them back to this through the API. This will be automated. However, I do just want to see how it's going to sound right here.

8:26 >> Good. And we can see the the one thing that's not necessarily translatable. My name is just spoken as >> Oh, all right. Let's try the other one. >> It's funny. Take two almost sounds to me

8:44 more like how I would probably sound if I had an elementary understanding of how to speak Chinese because the pronunciation sounds more like Englishified. Take one sounded like

8:54 somebody who had a more proficient grasp of Chinese. Now, I'm actually not going to pick which one is better here because either of these would be totally fine. And

9:09 really, the whole point of this is just knowing it is possible to have my voice cloned, but make me speak a different language that I can't actually really speak in real life. So, that will allow

9:19 me to get closer to what we saw right here. So, deep fabula. >> That one sounds more like the take two where it just makes me sound like very

9:29 basic Americanized. >> Hey, what you need? >> Actually, that's way better. So, that's also part of the problem is whatever they're using to clone here, I would not

9:37 imagine is tuned for quality, but rather for just large batch jobs just to get the content out that's being stolen. So, with that, our next step now is essentially to move into one of the next

9:47 phases. But before we do that, I would like to have a little bit of fun here. So something I'd like to specifically just play with even if just for a few minutes is the emotion control. So in

9:57 the docs here we can see this is one of the options for text to speech and emotion control is essentially just consisting of a bunch of emotion tags that will change the voice to reflect

10:07 that specific emotion. So we can see there are 24 basic emotions that we can choose from right here. So there are also advanced emotions and these are ones that could perhaps lead to more

10:18 humorous results. And in that case, I would like to just try a couple of these with some basic speech with the clone voice we had right there. Now, when I was looking through these docs, I did

10:28 see that it's probably best to have at least a minute of clean audio. So, I may end up reconing the voice for the actual like deployment of the stack we're building here. But for now, I just

10:37 wanted to see how it was with a short 15-second clip. And I think some of these will make their way into it and be kind of funny. And we can also include some if we just click on the tags right

10:46 here. And there's the option to auto tag as well, but that would not be as entertaining. So, let's just see how this one sounds. All right. And then

11:14 we'll try a a clear throat moaning an audience laugh. I don't know if you can stack them like this, but we'll find out. I should have skipped the first laughing. Let's just

11:40 >> I think it may have sometimes that happens where yeah in this quant 5090 for these simple D dollars gone push just to see how this performs in this quant.

12:02 Let's see if take two maybe was a little more. Give me the RTX5090 for the MSRP of $2,000. >> Good. Just want to see real quick for a

12:21 second. Give me the RTX5090 for the MSRP of $2,000. Oh, give me the RTX5090 for the Give me the RTX5090 for the MSRP of $2,000.

12:36 >> Sounds like my grandpa. >> Hey guys, and welcome back to the channel. Today we are going to be testing Fish Audio S2.1.

12:52 Hey guys and welcome back to the channel. Today >> I put that in the intro. >> Hey guys and welcome back to the

13:01 channel. Today we are going to be testing Fish Audio S2.1. All right, let's just try with both English. Today we are going to be testing Fish

13:15 Audio S2.1. Now I'm going to be doing one more voice clone. However, this is about a 70-second or right there. Oh, a minute and 18 seconds of like my voice and this is just pulled from that same

13:26 Quen 3.8 next video. And because they do mention a lot of times here on the platform that if you have a voice that's like around a minute and things like that, it's definitely the one we had

13:35 before was kind of in the orange to yellow portion. This is deep in the green, which means good enough. So, this should definitely give us a higher quality output, especially as we're

13:44 going to want some longer speech generation. So, I'll have this analyze it and then I'll basically do one quick test with it and then we'll move into actually figuring out how to pipe this

13:54 into a workflow that uses like a local model to transcribe the videos and then slice them based off of timestamps and things like that. Now, I do want to just quickly check it. So, I'm going to add

14:05 another voice right here to our new technical explainer voice and then we'll be able to just test it real quick. Hey guys, and welcome back to the channel. We had the speed set too low. That was

14:17 my mistake. All right. >> Hey guys and welcome back to the channel. >> That's very good. And I want Hey guys and welcome back to the

14:27 channel. And I wanted to hear it both in Chinese and English just cuz it's harder for me to judge if it's a good clone or not when I'm speaking Chinese. Unless it just starts out like this one did, and

14:35 it's not. All right. So, I'm going to be using GPT56 Soul on extra high right here just through the chat GPT Mac app. And I have written out a somewhat intricate prompt just going over the

14:47 overall goal for exactly what we're doing right here. I suppose I will read this back just verbatim. So I've said I need to perform the following build. A video is placed on the host system 30

14:57 minutes average length and then gets sliced up into sections based on timestamps provided with the video. Then the timestamp sections are fed to a local model to transcribe the speech

15:07 into text creating a transcript of that section of the video. The transcript will be translated to Chinese and the translated transcript will be sent to the fish audio API to recreate the audio

15:18 for that video section by using my clone voice in Chinese. Then the audio snippets will be placed back into the video replacing the original English audio. The video will be stitched back

15:28 together into the fulllength product and then ready to be uploaded to a platform with the audio now my voice cloned into Chinese. We will also need to be sure that the cloned audio lines up more or

15:38 less with the source audio in terms of temporal placement in the video. Perhaps some form of waveform/veillance analysis of the source clips audio that then gets compared to the clone clips audio. So we

15:50 can autonomously nudge the audio left or right to make these match basically. So like if my mouth is moving and then there's no audio playing, that's like what that section tries to fix just to

16:00 ensure the audio is more or less like at the same time as the source. Um, if that makes sense. I've said I believe the Quen 3.8 next model has the capacity to ingest these slices of videos and

16:11 transcribe them. So, we can use this for the initial test build, ask any follow-up questions needed, and set the environment up for this on the 7 machine, which is the 6000 Pro machine

16:20 behind us. So, I'm going to press enter here. Now, I may off the top of my head, it may be some overkill to actually be using a full-on like larger LLM to just transcribe sections of these videos. It

16:32 is probably a bit easier to basically have a system that takes the video, separates the audio, slices the audio only up based on timestamps and then feeds it to a light transcription model

16:43 to turn it into like a transcription that then gets translated. However, I just want to see what this will make essentially because this is probably a more overpowered way to do this. That's

16:54 not Yes. Okay. Has no audio input. Good. So basically, it's saying, "Hey, this model doesn't actually take audio in. So I'm going to set this build up with a dedicated speech pipeline and then using

17:05 Quen only to actually translate the transcribed speech into Chinese, which will probably be better." Oh, cool. It actually figured it out before I'd even noticed that the Blackwell machine host

17:15 had changed to 20 and not 7, which is correct. Cool. And it's saying for synchronization, I'm going to I'm not going to correlate the English and Chinese waveforms directly. The

17:25 phonetics make that signal misleading. The pipeline will instead match source speech onset, offset, and duration. Then use fish speed control plus a small final time stretch. It will lay the

17:34 rebuilt audio onto the original timeline and stream copy the video, avoiding cumulative stretch drift stitch drift. Good. Okay. And this is why I like just giving it an overview of my approach to

17:44 this because it will point out a lot of places like, hey, this is wrong. You're wrong here. So, I'm going to suggest this instead. And depending on the strength of the model you're using for

17:53 this, that can be a good thing or a bad thing. So 56 Soul has completed the validation just based off of the stack that it put on the system and ensuring that it does properly run on that

18:03 specific system. Next up, a couple of things. One of which I can't necessarily show on camera because I do need to get my API key and then the voice ID and things of that sort. So I am going to

18:12 create one right here. And I'm not going to be able to show this obviously at least past creating it. Then next up, I'm going to get the voice ID as well for the one we're going to use. And this

18:21 is the one that was a minute and 18 seconds of clone voice and not the 15-second one. Although they were both I mean the 15-second one quality-wise really didn't seem that far, but maybe

18:30 over much longer generations the difference will become a bit more apparent. Next up, I'm giving it one additional thing that I'd like, which is to essentially wrap the stack that it

18:39 just made into a web UI. So that way that machine will be accessible from this one and it will simplify the process of actually choosing a source clip like putting in the API key and

18:48 everything like that and just getting everything set up without having to do any like terminal work. All right, we have our front end finished and I did also just tell it to keep the quen model

18:56 loaded in memory because it was going to load and unload depending on when it was being used which we don't need for this. It will just introduce a bit more latency to the endto-end process. Now,

19:05 next thing I need to do here is go into settings and enter in the fish audio API key as well as the clone voice ID. We also do have some other metrics right here just pulled from the actual fish

19:15 documentation in terms of like how to use it this thing that we had pasted to it. So, I am going to now give it the things that are necessary like API key and voice ID which I can't show. All

19:24 right, I have entered the fish credentials in here now and we do also have the correct S2.1 Pro model selected. So, let's test the saved credentials here. Save it. Oh.

19:35 Now we'll test the credentials. Fish connected technical explainer voice S2.1 Pro. Perfect. So that is the correct model S2.1 Pro. And then the technical explainer voice is the minute

19:47 and 18 second clone version of the voice. Good. So first and foremost, we have this all set. Automatic section length. So this is when no timestamps are supplied in terms of actually

19:56 chopping up the video, which may actually be something that works. All right. So after a bit of having GPT56 soul troubleshoot some of the issues that we were having here basically

20:05 Chrome didn't have access to the local network which was causing it to not be reachable from Chrome and only from within the chat GPT app. Then additionally to that we had a CUDA

20:14 dependency issue for the local speechtoext that was causing a problem. So now we should hypothetically be good to go. I do want to ensure that in the settings here. Okay. So, I still need to

20:26 Good. My API key is selected and as we saw earlier, the voice when we tested the saved credentials was correct and it is using the S2.1 Pro model. So, we're basically good to go now. And let's just

20:38 try this with a much shorter source video as like I hate to use AI lingo, but let's run a quick smoke test just to make sure it works. So, this is a 26 GB file. I'm going to pass on including any

20:51 timestamps for this right now and we'll just see how this goes with the start Mandarin dub. All right, so we have our old previously failed job there that was just um has

21:01 been fixed now. Okay, so let's see if we get past that 65%. And the computer is using the local Quen model to translate the transcription of the speech into Chinese. So then it will send that

21:13 Chinese transcription to Fish Audio to S2.1 Pro to generate with my voice cloned into Chinese and then put it back on the video and things like this. Okay, apparently and this is actually quite

21:25 cool. So we can see this autosliced this video. We have the transcriptions actually into Chinese right here. And these are what are going to be sent to Fish Audio to actually generate the

21:34 speech. So hypothetically now, sorry I'm talking louder or faster cuz I got a little excited because this is kind of cool. Like this is showing a combination of like a local and a cloud service

21:45 pipeline to automate something that is kind of a bit of a task to take my video, transcribe it into Chinese, put the audio back over the video with the clone voice. I mean, this is a pretty

21:55 cool thing that was just essentially vibe coded together in like not even a whole afternoon. All right, the download has just finished. Let's see.

22:09 >> That's awesome. That is very awesome. >> And this sounds like me, too. Like, well, it is cuz it's my clone voice. But check this out. I mean, this is like

22:28 this is a sick pipeline. Build skills. I like this. All right. So, here's the thing. That was like our validation test. Now, let's actually do this. Let's

22:38 slice and dice the full-on Quen 3.8 video, which is I think between 30 or 40 minutes. All right, we now have the timestamps for the full length video, which is probably around 33 minutes now

22:48 that I'm seeing this. So, we're going to just paste those here into this web UI, and we're going to upload the source video that is the full length one. We can see right here, this is a 6.6 GB

22:58 file. I'm not sure how this is going to handle actually compressing it and things like that. We already do have the fish audio S2.1 clone voice selected in the API and everything like that. So

23:09 assuming that this actually can handle ingesting a video of this length, keeping in mind it is chopping it up. This is our first actual like end to end smoke test as Claude would say. All

23:19 right, it has just completed uploading. Now that took quite a bit of time. So if we scroll up, this is our current job as it was a six and a half gigabyte video that we needed to get done. So, it's

23:29 transcribing locally and I can actually hear the GPU working right now. Yep. So, it is I don't actually know how long it will take, but it's using Quen 3.8 next at a Q4KM quant. Okay, that seemed

23:40 actually pretty decently quick. Uh, now it's using the GPU and the model because now it's translating the transcriptions to simple Mandrin. So, that's what's using the Quen next Quen 3.8 next at a

23:53 Q4KM quant. And then once the translation has been complete, I believe we'll see that table again showing the input source text and then the translated text. And then the translated

24:03 text will be sent to Fish Audio for the S2.1 Pro model to translate using the clone voice into Chinese. And then it will give us the video back once it's kind of spliced everything back

24:14 together. So, it's currently generating clone speech and it has been for a while. But fortunately, something that was included in this build is this specific metric

24:23 right here of how much free space is on the system that's running this. And that's been actually going down like pretty decently. So that shows us that it is actually pulling things, making

24:33 new files, and creating this video for us. And just like that, we received all of our speech. So it's fitting and remixing the audio. Now, we can see here that we

24:44 have all of the transcriptions. And then the translations were done by the local Quen model. And the transcriptions were done locally as well. And the slicing just happened based off of the

24:54 timestamps. And then each of these individual snippets would have been sent to the Fish Audio S2.1 Pro endpoint with our clone voice and then brought back and then put back together for like this

25:07 video. Okay. And it says it's all done now. So we can just download it. That'll probably take a little bit of time and then we'll be able to actually see how this handled a full 33 or so minute

25:18 video entirely translated to Chinese in my voice. Our transcribed video download has completed. So let's take a look at this and see. Keeping in mind this is a 32minute video.

25:43 I screwed something up, didn't I? All right, so this is pretty interesting. We've received the full downloaded clip and >> see how it's like

25:53 >> it's too slow. >> They're like it's better in some places. So I asked codeex essentially like what exactly has gone on here and it's interesting because that was actual

26:09 behavior of our pipeline. So we can see right here this is intentionally slowed down by the current fitting logic. So it asked fish audio to actually slow down some of these clips because it does have

26:20 that capability to either speed up or slow down the generations as we saw earlier. So, it's essentially trying to fit the speech into the same time span as the English version, which is not the

26:32 case. It would have been spoken faster when translated into Chinese. So, that's why it seems slowed down because it's trying to make it fit the original length, even if it's not like a one one.

26:42 So, it's just interesting to see stuff like this that pops up that's not even something you would have probably thought of prior to do this. So, the proper fix is to treat those timestamps

26:51 as chapter boundaries. automatically subdivide each chapter using whisper speech timing into roughly 10 to 30 seconds. Synthesize each near onetime speed and position it at its original

27:01 speech onset. Shorter Mandarin should leave natural gaps rather than being stretched across an entire chapter, which is what's happening right now. So, I'm going to say do this and we'll have

27:12 to wait again for a new render, but that will give us a more accurate representation of the proper way to do this, keeping in mind there are differences between translating. And

27:21 that's why this clip right here felt so much more natural >> proin DDRC. So that's actually kind of interesting and something that would have only

27:34 reared its head doing a full length test just like this where it had to make these shorter snippets basically fit the length of the full section. So I'm going to have this change it and we'll try it

27:44 again just with the more optimized settings. And apparently 56 soul is going to regenerate the existing job without another upload. So we may not have to actually do anything in the web

27:53 UI for this specific task. So this should hypothetically be the version that is okay. >> This basically is exactly what we wanted.

28:10 >> So the I think the whole takeaway is the voice definitely 100% sounds like me speaking Chinese. it seems like and something I didn't even really I guess appreciate from those content thiefs is

28:21 the fact that it actually takes a little bit of like fine-tuning in like art to actually get it to properly sync up with a video that was created in English and then swapping it into Chinese. It's

28:32 almost going to be interesting to figure out exactly how to fine-tune this system in a way that properly stretches the audio out over the video without keeping too many gaps where it's basically just

28:43 like with no sound playing or something of the sort. However, the core component of this is actually getting a properly translated voice of myself speaking Chinese, something that can

28:55 programmatically be used like we've done here through the API. So that's going to be something interesting. But really like the whole point of this again is the fact that we

29:13 have a programmatic solution that will actually just hit the fish audio API right here with this local studio. And this is more or less automated to the point where I just upload the video to

29:24 it and because of what we find out about found out about the timestamps causing the audio to slow down artificially. Actually not using those seems to be a better metric. So it's basically like a

29:34 system now where the source video gets uploaded here. It gets transcribed and then the local Quen model translates the video's transcription into Chinese and then it sends that Chinese text to the

29:45 fish speech API here using our clone voice. I believe it was the technical explainer voice. That was the one that was the longer minute and 18 second clip. And then it sends it back here to

29:56 the local system, stitches it back together in the portions of the video that the clips correlate to, and then exports the video, but with the language changed from English to Chinese using my

30:07 clone voice. This was almost like a little vibe build basically of an actual like functional thing that I have been needing because I would like to be able to upload my own content to that

30:17 website, especially since it's being uploaded by so many different people with so many different approaches. Some just use subtitles, some clone my voice and make it seem like this. So, I want

30:27 to kind of get on top of that. And this was basically the first step in it was just building out a pipeline that can be autonomous and like replicatable and something that uses both cloud for

30:38 something a little more potent in terms of like the clone voice which is why we're using S2.1 Pro and then basically having some of it offloaded to to a local process like being able to

30:48 translate the transcription into Chinese and then just automate the API and things of that sort. So that was like a pretty interesting build. I think I had a lot of fun with it and obviously just

31:00 playing with this in the like voice studio as well is a ton of fun. So I think this is something that I now basically have the groundwork laid in terms of a system that will get me to

31:09 where I can almost automate uploading my content onto Billy Billy as well natively in Chinese without having to do much work in terms of like manually dealing with any of this. So, that is

31:20 probably going to conclude today's video, which was partially a test of Fish Audio's S2.1 Pro model, which is very, very good as we see here. Even with the like 15-second voice clip, it

31:32 cloned it very, very well. So, this was really cool. And then also basically like vibe coding something that integrates the API for fish audio into a system that runs on a local machine. So,

31:43 this was um turn one English master into a Mandarin delivery. That's probably a better way than I could have summarized what happened here. So, if you have any questions, please feel free to leave

31:53 them in the comments. Again, thanks to Fish Audio for letting me play with S2.1 Pro and using it through the API and automating and having a little fun, too. So, if you have any questions, please

32:03 feel free to leave them in the comments. And thanks for watching.