Show HN: Open source framework OpenAI uses for Advanced Voice
1 year ago (github.com)
Hey HN, we've been working with OpenAI for the past few months on the new Realtime API.
The goal is to give everyone access to the same stack that underpins Advanced Voice in the ChatGPT app.
Under the hood it works like this: - A user's speech is captured by a LiveKit client SDK in the ChatGPT app - Their speech is streamed using WebRTC to OpenAI’s voice agent - The agent relays the speech prompt over websocket to GPT-4o - GPT-4o runs inference and streams speech packets (over websocket) back to the agent - The agent relays generated speech using WebRTC back to the user’s device
The Realtime API that OpenAI launched is the websocket interface to GPT-4o. This backend framework covers the voice agent portion. Besides having additional logic like function calling, the agent fundamentally proxies WebRTC to websocket.
The reason for this is because websocket isn’t the best choice for client-server communication. The vast majority of packet loss occurs between a server and client device and websocket doesn’t provide programmatic control or intervention in lossy network environments like WiFi or cellular. Packet loss leads to higher latency and choppy or garbled audio.
Imagine being able to tell an app to call the IRS during the day, endure the on-hold wait times, then ask the question to the IRS rep and log the answer. Then deliver the answer when you get home.
Or, have the app call a pharmacy every month to refill prescriptions. For some drugs, the pharmacy requires a manual phone call to refill which gets very annoying.
So many use cases for this.
As costs of humanlike communications decrease, so will Sybil attacks and spam.
The IRS is notorious for resistance to tech change, don't be surprised if they unplug the phones and force you to walk in to ask your question.
What is the value add here? Save sometime for technocrats and technoadjacents for a whole of 3 years before victims of spam adapt?
Also this has been solved already just mail your question like the rest of mortals.
It would be really nice if the IRS would ALLOW you to walk in and ask a question!
5 replies →
They'll just put up "captchas" or whatever.
The point if phone lines is to waste the client's time. Not to have the client waste their time.
We do this exact thing at getvibrato.com. You can schedule calls like these, or even do more advanced automation with Zapier.
This is really helpful, thanks!
OpenAI hired the ex fractional CTO of LiveKit, who created Pion, a popular WebRTC library/tool.
I'd expect OpenAI to migrate off of LiveKit within 6 months. LiveKit is too expensive. Also, WebRTC is hard, and OpenAI now being a less open company will want to keep improvements to itself.
Not affiliated with any competitors, but I did work at a PaaS company similar to LiveKit but used Websockets instead.
> LiveKit is too expensive
Most of it is open source, especially the clients, although they do feel quite ad hoc hacked together (a possible side effect of WebRTC evolution).
Would totally agree on OpenAI moving away. The description of the agent here sounds like a big hack just to get around the fact temporarily the model server expects audio over sockets instead.
Which components feel ad hoc?
In most real applications, the agent has additional logic (function calling, RAG, etc) than simply relaying a stream to the model server. In those cases, you want it to be a separate service/component that can be independently scaled.
4 replies →
Field CTO — hi @Sean-Der :wave:
Fractional CTO sounds like a disaster lol
My bad, he was Field CTO.
Super cool! Didn't realize OpenAI is just using LiveKit.
Does the pricing breakdown to be the same as having a OpenAI Advanced Voice socket open the whole time? It's like $9/hr!
It would be theoretically cheaper to use this without keeping the advanced voice socket open the whole time and just use the GPT4o streaming service [1] for whenever inference is needed (pay per token) and use livekits other components to do the rest (TTS, VAD etc.).
What's the trade off here?
[1]: https://platform.openai.com/docs/api-reference/streaming
Currently it does: all audio is sent to the model.
However, we are working on turn detection within the framework, so you won't have to send silence to the model when the user isn't talking. It's a fairly straight forward path to cutting down the cost by ~50%.
Working on this for an internal tool - detecting no speech has been a PITA so far. Interested to see how you go with this.
3 replies →
Can I currently put a VAD module in the pipeline and only send audio when there is an active conversation? Feel like just that would solve the problem?
You dont get charged per hour with the openai realtime api, only for tokens from detected speech and response
That’s some crazy marketing for a „our library happened to support this relatively simple use case“ situation. Impressive!
By the way: The cerebras voice demo also uses LiveKit for this: https://cerebras.vercel.app/
There’s a ton of complexity under the “relatively simple use case” when you get to a global, 200M+ user scale.
80% of the times I’m experiencing choppy audio on my iPhone 15 Pro Max (18.1b) on Voice Mode (Standard and Advanced). My internet connection is FTTH and WiFi 7 state of the art router.
I wonder if this is because bugs or the crazy load livekit may be going through given the popularity in ChatGPT voice modes right now.
1 reply →
Olivier, Michelle, and Romain gave you guys a shoutout like 3 times in our DevDay recap podcast if you need more testimonial quotes :) https://www.latent.space/p/devday-2024
I had no idea! <3 Thank you for sharing this, made my weekend.
You guys are honestly the best
Is there anyone besides OpenAI working on a speech to speech model? I find it incredibly useful and it's the sole reason that I pay for their service but I do find it very limited. I'd be interested to know if any other groups are doing research on voice models.
Yes. Kyutai released an opened model called moshi : https://github.com/kyutai-labs/moshi
There's also llama-omni and a few others. None of them are even close to 4o from an LLM standpoint. But moshi is called a "foundational" model and U'm hopeful it will be enhanced. Also there's not yet support for those on most backends like llamacpp / ollama etc. So I'd say we're in a trough but we'll get there.
There’s Ultravox as well (from one of the creators of WebRTC): https://github.com/fixie-ai/ultravox
Their model builds a speech-to-speech layer into Llama. Last I checked they have the audio-in part working and they’re working on the audio-out piece.
When I asked advanced voice mode it said that it receives input as audio and generates text as output.
It is mistaken because it has no particular insight into its own implementation. In fact the whole point is that it directly consumes and produces audio tokens with no text. That's why it's able to sing, make noises, do accents, and so on.
I wonder when Azure OpenAI will get this.
I'm working on a PR now :)
This suggests that the AI "brain" receives the user input as text prompt (agent relays the speech prompt to GPT-4o) and generates audio as output (GPT-4o streams speech packets back to the agent).
But when I asked advanced voice mode it said the exact opposite. That it receives input as audio and generates text as output.
Both input and output are audio. This post is about bridging WebRTC audio I/O with an API that itself operates on simple TCP socket streams of raw PCM. For reliability and efficiency you want end users to connect with compressed loss-tolerant Zoom-style streams, and that goes through a middleman which relays to the model API.
Who did you ask? ChatGPT? Not sure if you understand LLMs but its knowledge is based on the training data, it can't reason about itself, it can only hallucinate in this case, sometimes correctly, most times incorrectly.
This is also true for petty much all humans and bypassing this limitation is called enlightenment/self realization.
LLMs don't even have a self so it can never be realized. Just the ego alone exists.
6 replies →
Nice they have many partners on this. I see Azure as well.
There is a common consensus that the new Realtime API is not actually using the same Advanced Voice model / engine - or however it works - since at least the TTS part doesn’t seem to be as capable as the one shipped with the official OpenAI app.
Any idea on this?
Source: https://github.com/openai/openai-realtime-api-beta/issues/2
It's using the same model/engine. I don't have knowledge of the internals, but a different subsystem/set of dedicated resources though for API traffic versus first-party apps.
One thing to note is there is no separate TTS-phase here, it's happening internally within GPT-4o, in the Realtime API and Advanced Voice.
Thanks
so the WebRTC helps with the unreliable network between the mobile clients and the server side. if the application is backend only, would it make sense to use WebRTC or should I go directly to realtime api?
That was cool, but got up to $1 usage real quick
We had our playground (https://playground.livekit.io) up for a few days using our key. Def racked up a $$$$ bill!
How much is it per minute of talking?
3 replies →