Name: Sindarin
Uploaded: 2024-12-16T11:21:30.509Z
Duration: 2 min 33 s
Description: An AI-powered platform enabling developers to build and deploy advanced conversational speech agents with ultra-low latency and human-like interactions.

sindarin

An AI-powered platform enabling developers to build and deploy advanced conversational speech agents with ultra-low latency and human-like interactions.

Sindarin

I found a voice AI platform that's a 4 second quicker than vapy in response time. For some reason, nobody is talking about it. You might think a second, that's not that much. But in voice AI, that's the difference between a clunky conversation and one that flows naturally. This new platform isn't just faster; it's a new minimum. Once you experience it, you genuinely will not be able to go back to fappy. The improvement is that dramatic. 

Let me show you a side by side comparison that demonstrates why this is such a game changer. Can you tell me the time? Yes, but first can I get your name? Yes, but first can I get your name? 

Going to be honest, there are still two major problems with voice AI. These issues are holding back the technology from reaching its full potential. The first is the issue of latency. To have a humanlike conversation, response time should be able to be almost instantaneous, around 300 to 500 milliseconds. This speed of response is crucial to maintain a natural flow of dialogue. Sometimes in conversations, we wait two or even 3 seconds before responding. But those types of pauses are used for effect and are intentional. They convey meaning, like emphasis or contemplation. The key is that these pauses are controlled by us. 

However, most current voice AI platforms out there average 2 to 3 second response time. This means that even when responses should be quick, the AI is unable to respond at the speed needed. This consistent delay isn't the end of the world, especially over the phone. However, if we are striving for a natural humanlike conversation, reducing that latency would be huge. 

The other issue we have is bad turn-taking technology. This is another crucial aspect of natural conversation that most voice AI platforms struggle with. What I mean by turn-taking is the ability to manage a conversation. This includes knowing when to start speaking, when to pause, and how to handle interruptions. In human conversations, we do this effortlessly and often without even thinking about it. 

Now here's why knowing about this platform gives you such an advantage. These two major issues I just mentioned are significant.

Mentioned are at the forefront of what they are trying to solve. First, the latency improvement is remarkable. Response times are down to around 1 to 1.5 seconds. While not perfect, it's still a huge leap from the 2 to 3 seconds we are used to. I promise you, once you experience this reduced delay, other voice AI platforms will feel frustratingly slow. 

But that's not all. Their turn-taking technology is also very good. The AI handles interruptions very naturally. It can stop mid-sentence just like a human would. Okay, so I'm sure you're all eager to find out what this platform is. I stumbled upon Cerin about a month ago when reading a feedback post on Vapy forums. Someone mentioned that Syn's turn-taking technology was superior to Vapy, and as this is my field of work, I decided to take a deeper look. 

I'm going to be honest. When I first read the post, I was a little skeptical. I thought I haven't even heard of Cerin. There's no way it can be that much better than Vapy. But I was wrong. I quickly built a test agent, and the moment I started talking with it, I was genuinely blown away. The improvement in latency was like immediately noticeable. It felt almost instantaneous. 

Now, I will say that Cerin's UI and functionalities are a little bit lackluster compared to Vapy. However, the latency is so good that it more than makes up for that. I can really only talk so much about how much better the latency is without showing you. So what I'll do is run a challenge between Vapy and Cerin, and then I will show you how to build your Cerin. 

Essentially, what I did for this test was build a program that is able to measure the voice AI latency down to the millisecond. For the technical people who are interested, here is how it works. You start by inputting an audio recording of you talking with your voice AI, and then you press run. 

What happens in the back end? You first input the audio recording of you talking with your voice AI to the program, and then what it will do is isolate the voices and remove any sort of noise. This is basically to clean up the audio. Then a program runs through that.

Entire audio and finds all the moments of silence within it. Then what we do is transcribe that conversation and find their two speakers and what parts of the conversation they were speaking. Now how we actually measure the latency is when I finish speaking. So let's say I'm speaker two, the human. How long is the silence between me finishing speaking and when speaker one starts to speak? That being the A, and that's how we calculate the latency. 

Then what we do is do this for every time the conversation was turned. So when I stopped speaking, the AI did again, and then we average out all those latencies in order to get the average for that conversation. I'm not sure that explanation was that good, but I promise you the program works. 

Hello, this is Lola. How can I help you? Hi there, can you tell me the time? Yes, but first can I get your name? It's Hugo. This will just take a sec. Hugo, the time is 15:34. Perfect, thank you. You're welcome, Hugo. Have a great day. You too, bye-bye. 

Hello, this is Lola. How can I help you? Hi there, can you tell me the time? Yes, but first can I get your name? It's Hugo. Checking the time for you now, Hugo. Hugo, the time is 15:25. Perfect, thank you. You're welcome, Hugo. Is there anything else I can help you with? No, goodbye. 

Okay, so this little tutorial right here will be mostly for people who have built some sort of a voice AI system on a different platform before and have an understanding of what voice AI is and the kind of main components that are needed to build voice functionality. So to start off, we're just going to go to Cinder in.

Tech which is their landing page and we're just going to press get started. That's going to take you to an authentication page where you can sign up. Once you sign up, it's then going to take you to this dashboard right here. As you can see, you have person owners, transcripts, API docs, billing, settings, logout, etc. The different sections are self-explanatory. Transcripts shows your transcripts. API documentation is your API documentation. Billing is your billing, etc. 

I'm going to focus on personas because that's kind of where you create your assistant. Where you choose what voice model you're using, what prompt you have, your actions or tools. So we'll start off by pressing create persona, naming it whatever we want. So I'm going to name it Lily and the LLM we will leave as metal Lama 3.1 70 billion as that's quite a good model that's in between speed and reasoning. 

Now here you would press create, but I'm not going to because I've already created the agent, as you can see here. So once you get to this part of it, this is where you kind of edit and actually do all the stuff with your assistant. We start off in the general tab where you have your initial message. So this is basically the first message that your assistant will say when starting a conversation. So mine would be, "Hello, my name is Lily. How can I help you?" This is never going to change no matter what your conversation looks like because your agent will always say the first message. 

Then you have your prompt. So this is again self-explanatory. If you've used any other platform like Fappy, you'll know exactly what a prompt is and what it does. Then you also have the prompt again. This is really self-explanatory if you've built any other voice AI assistant on another platform like Vapy, let's say. But there are a few different things when it comes to cindering that you need to learn. 

Down here, as you can see, we have prompt variables, current date, time details, action schema, and scenarios. These are basically variables that you need to put in your prompt if you want your persona to have the ability to do things. So action.

Schema. This is basically your tools. This is what allows your persona or your assistant to do things and manipulate things outside of just a prompt and saying things. So that's always useful. As you can see in my prompt, I have the action schema there. Current date and time. I wouldn't recommend using it because it just tells you the date and time but for one specific time zone. So the time for someone in London might be different. But also just remember that most agents wouldn't need the date and time so that doesn't matter too much.

Then you also have details which actually I don't know what it is and what it means. But I will ask the founder of Cerin as I do have contact with him. Also, if you have any questions about Cerin, anything whatsoever, things that even I can't answer, drop them in the comments because I'm going to link this video to the founder and tell him that if people have questions and you know people who are new to the platform have questions, um, who are watching this video will be commenting and therefore he can answer your comments even if I couldn't.

We also have scenarios which is basically like their RAG feature. As you can see here, you can add context and basically give different scenarios and what the persona, this is what the assistants are called in Cerin, it's called a little bit different. Basically, you can tell the persona what to do when something happens. Now I usually just create like a knowledge base SL RAG action which basically allows the agent to do the normal knowledge-based search and I just build RAG technology in a make.com scenario.

Now, like I said, actions is the tools alternative. So let me give you a bit of an explanation of that. As you can see here is our actions. So we can give it a name. So I named mine check time. The description, this is like the kind of prompt that you have with any sort of tool. I'm going to just actually change this to this action fetches the current time. As you know, it's not called tools. I'm quite used to the Vapy terminology. Then you can add some examples. So if a user says my time zone is BST because I'm from.

London then make sure to use this action and the reason why I have it as time zone is if we look at the prompt I say number three. If the user wants to know the time, ask for what time zone they are in. Then check and provide the current time. So when using this action, the user will always say before what their time zone is. Then you also would add the action. Then you would also add in the example what parameter you are adding, which would be timeone BST. This again helps the Persona to know how it should structure the parameters it is sending to the web hook. 

Now you're going to go into advanced settings over here and go down to action type and basically turn this on and change it from update state to web hook API call. Then you're going to paste in your make.com URL, which will be your um web hook URL. Chuck it in there and then make the method post. Here you can add your parameters. So in the body, you can make it something like name and then value the user's name, which is just your kind of prompt of what it is. So as you can see there, I have the time zone and the users time zone and that's really it. That's all you have to add for your actions. 

In order to do that, you would just press add new and then do custom action. I just went over and as you saw, I went over a pre-existing action. But to create a new one, you would just press add new and do custom. As you can see, you have this full JSON schema over here and that's basically what's in the prompt and it's encapsulated in this variable over here. So that basically allows your agent to call actions. 

I am going to skip over scenario because personally I don't think it's that valuable and you can just build a rag SL knowledgebase tool or action and allow it to do it that way. Now for the voice, I just chose this one over here 656 F6 because I believe her voice is very good. It's 11 Labs. Let me play it for you. Hello and welcome to Persona and yeah, and also this voice was really quick and it's 11 Labs so I really liked the voice. 

Next, we can go to LLM. Like I said, I chose the default which was 3170 billion again fast and has a good.

Reasoning so I'm going to leave it to that. But I will change the temperature to 0.4 just because I think having the temperature lower and it listening to my prompts better is overall better for the Persona. 

Then we also have interjections which is basically allowing the agent to kind of reprompt or resay things when there is silence. So let's say it asks me something and I don't say anything for 10 seconds. It will say hello are you still there something like that. And I'm just going to leave this at the because I think that's a good starting point. 

Then life cycle events I don't really know what this is so I'm not even going to get to it. I think it's something to do with after a conversation you can send something to a URL or like for example the transcript things like that. But the API documentation doesn't really talk about that so I'm going to leave it for now. 

So yeah now we can just test this agent. I'm going to test it and see if it is able to run this tool. So I'm going to just run it and we're going to press start. 

Hello my name is Lily how can I help you today? Hi there can you tell me the time? Before I can tell you the time may I know your name please? Yeah my name is Hugo nice to meet you Hugo. Do you want to know the current time? Yes please. 

Hugo what time zone are you in? Um I'm in BST because I'm from London. Hugo just a moment while I check the time for you. Hugo the current time in BST is 1553. 

Okay so as you see it was able to check the time and also the time for me is 1553 as you can see. So it was able to put in that time zone and change the time to the correct time zone. And as you saw that level of latency was so so low it was almost instantaneous. So much quicker than vapy. 

Now once you have your persona and you are happy with it you can deploy it by essentially deploying it either via code via Twilio or Unity. So Unity is for game development if you want your NPC to have you know voice AI features then you would implement it that way. 

Now if I were to deploy this I would integrate it with Twilio. This is like you know the normal standard like you would in.

Vapy and here you have a six-step guide on how to deploy it. Again, it's very simple but I'm not going to go too in depth in this because I've actually never deployed and I'm sure that these instructions right here are good enough for you to be able to do them. So yeah, that's really it when it comes to cindering. 

I didn't really go too in depth when it comes to the API documentation side of things as that's something for you guys to learn yourself and I don't know too much about cerin myself. I've only known about it for a month and therefore I do still have some holes in my knowledge about it. 

Like I said, it is kind of a shorter tutorial just to show how the UI looks, you know how the action schema works for example. You know, that might be something that most people wouldn't have understood if I didn't show you. 

And yeah, like I said, if you have questions about cindering just drop it in the comments and I'm sure I can get the founder to answer these questions, which is the best type of support you can possibly get. Also, let me know in the comments if you want to see a video where I compare all the different voice AI platforms out there and their latency. 

Like I said, I built a program to measure latency of voice AI platforms, you know, like the true latency because VY for example says oh yeah this will be 900 millisecond latency but then I actually talk to it and I find the latency is 2-3 seconds even. So if you want me to do that and compare all the platforms out there so you can see each platform ranked on a leaderboard, then let me know.

Note, that there may be slight differences in text and video. Watch video for the original content.

Sindarin Video