The first time I met Tavus, their engineers (incl Brian!) were perfectly willing to sit down and build their own better Infiniband to get more juice out of H100s. There is pretty much nobody working on latency and realtime at the level they are, Sparrow-1 would be an defining achievement for most startups but will just be one of dozens for Tavus :)
Maybe infiniband is a bit more than we can handle. That technology is incredible! You are right though, we have been willing to build things we needed that didn’t exist yet, or were not fast enough or natural enough. Sparrow-1, Raven-1, and Phoenix-4 are all examples that, and we have more on the way.
Hey @code_brian, would Tavus make the conversational audio model available outside of the PALs and video models? Seems like this could be a great use case for voice-only agents as well.
Awesome. We've been using Sparrow-0 in our platform since launch, and I'm excited to move to Sparrow-1 over the next few days. Our training and interview pre-screening products rely heavily on Tavus's AI avatars, and this upgrade (based on the video in your blog post) looks like it addresses some real pain points we've run into. Really nice work.
That’s great! I also built Sparrow-0, and Sparrow-1 was designed to address Sparrow-0’s shortcomings. 1 is a much better model, both in terms of responsiveness and patience.
I am always skeptical of benchmarks that show perfect scores, especially when they come from the company selling the product. It feels like everyone claims to have solved conversational timing these days. I guess we will see if it is actually any good.
Different industry, but our marketing guy once said "You know what this [perfect] metric means? We can never use it in marketing because it's not believable"
You should be skeptical, and try it out. I selected 28 long conversations for our evaluation set, all unseen audio. Every turn taking model makes tradeoffs, and I tried to make the best tradeoffs for each model by adjusting and tuning the implementations. I’m certainly not in a position as the creator of Sparrow to be totally objective. However we did use unaltered real conversational audio to evaluate. I tried to find examples that would challenge Sparrow-1 with lots of variation in speaker style across the conversations.
Btw while I think this is cool and useful for real time voice interfaces for the general populace, I wonder if for professional users (eg a dev coding by dictating all day), a simple push to talk is not always going to be superior, because you can make long pauses while you think about something, this would creep out a human, but the AI would wait patiently for your push to talk.
As a dev myself, I see a couple of modes of operation:
- push to talk
- long form conversation
- short form conversation
In both conversational approaches the AI can respond with simple acknowledgements. When prompted by the user the AI could go into longer discussions and explanations.
It might be nice for the AI to quickly confirm it hears me and for it to give me subtle queues that it’s listening: backchannels: “yeah”, and non-verbal: “mhmm”. So I can imagine having a developer assistant that feels more like working with another dev than working with a computer.
That being said, there is room for all modes, all at the same time, and at different times shifting between them. A lot of time I just don’t want to talk at all.
> Non-verbal cues are invisible to text: Transcription-based models discard sighs, throat-clearing, hesitation sounds, and other non-verbal vocalizations that carry critical conversational-flow information. Sparrow-1 hears what ASR ignores.
Could Sparrow instead be used to produce high quality transcription that incorporate non-verbal cues?
Or even, use Sparrow AND another existing transcription/ASR thing to augment the transcription with non-verbal cues
This is a very good idea. We currently have a model in our perception system (Raven-1) that performs this partially. It uses audio to understand tone and augment the transcription we send to the conversational LLM. That seems to have an impact on the conversational style of the replicas output, in a good way. We’re still evaluating that model and will post updates when we have better insights.
It sounds really cool, but I don't see any way of trying the model directly. I don't actually want a "Persona" or "Replica" - I just want to use the sparrow-one model. Is there any way to just make API calls to that model directly?
That’s unfortunate and certainly not what I spend my time dreaming about. My favorite use case for the elderly is as a sort of companion for sharing their story for future generations. One of our partners uses our technology to help elderly. But yeah, this kind of technology makes AI feel more natural, so we should be aware of that and make sure it’s used for good.
Literally no way to sign up to try. Put my email and password and it puts me into some wait list despite the video saying I could try the model today. That's what makes me mad about these kind of releases is that the marketing and the product don't talk together.
I tried talking to Claude today. What a nightmare. It constantly interrupts you. I don’t mind if Claude wants to spend ten seconds thinking about its reply, but at least let ME finish my thought. Without decent turn-taking, the AI seems impolite and it’s just an icky experience. I hope tech like this gets widely distributed soon because there are so many situations in which I would love to talk with a model. If only it worked.
Agreed. English is not my native language. And I do speak it well, it's just that sometimes I need a second to think mid-sentence. None of the live chat models out there handle this well. Claude just starts answering before I've even had the chance to finish a sentence.
Anthropic doesn't have any realtime multimodal audio models available, they just use STT and TTS models slapped on top of Claude. So they are currently the worst provider if you actually want to use voice communication.
I think Anthropic currently has a slight edge for coding, but this is changing constantly with every new model. For business applications, where tool calling and multi-modality matter a lot, OpenAI is and always has been superior. Only recently Google started to put some small dents in their moat. OpenAI also has the best platform, but less because it is good and more because Google and Anthropic are truly dismal in every regard when it comes to devx. I also feel like Google has accrued an edge in hard-core science, but that is just a personal feeling and I haven't seen any hard data on this yet.
I love Anthropic's models but their realtime voice is absolutely terrible. Every time I use it there is at least once that I curse at it for interrupting me.
My main use case for OpenAI/ChatGPT at this point is realtime voice chats.
OpenAI has done a pretty great job w/ realtime (their realtime API is pretty fantastic out of the box... not perfect, but pretty fantastic and dead simple setup). I can have what feels like a legitimate conversation with AI and it's downright magical feeling.
That said, the output is created by OpenAI models so it's... not my favorite.
I sometimes use ChatGPT realtime to think through/work through a problem/idea, have it create a detailed summary, then upload that summary to Claude to let 4.5 Opus rewrite/audit and come up with a better final output.
I use Claude Code for everything, and I love Anthropic's models. I don't know why, but it wasn't until reading this that I realized: I can use Sparrow-1 with Anthropic's models within CVI. Adding this to my todo list.
The one thing that really surprised me, the thing I learned that's affected my conversational abilities the most: turn taking in conversation is a negotiation: there are no set rules. There are protocols:
- bids
- holds / stays
- implications (semantic / prosodic)
But then the actual flow of the conversation is deeply semantic in the best conversations, and the rules are very much a "dance" or a negotiation between partners.
That's an interesting way to think about it, I like that.
It also implies that being the person who has something to say but is unable to get into the conversation due to following the conversational semantics is akin to going to a dance in your nice clothes but not being able to find a dance partner.
Yeah, I can relate to that. Maybe it's also because you are too shy to ask someone to dance. I think I learned that lesson: just ask, and be unafraid to fail. Things tend to work themselves out. Much of this is experimentation. I think our models need to be open to that: which is one cool thing about Sparrow-1: it's a meta-in-context learner. This means that when it try's and fails, or you try and fail, it learns at runtime to adapt.
The response timing in the chart in the blog post shows that even with perfect precision/recall Sparrow-1 also has the fastest true positive response times.
The turn taking models were evaluated in a controlled environment with no additional cascaded steps: LLM, TTS, Phx. This matters to get apples to apples comparison: without the rest of the pipeline variability influencing the measurements.
The video conversation examples are sparrow-1 within the full pipeline. These responses aren’t as fast as sparrow itself because the LLM, TTS, facial rendering, and network transport also take time. Without Sparrow-1 they would be slower. Sparrow-1 enables the responses being as fast as they are, and with a faster CVI pipeline configuration the responses can be as fast as 430ms in my testing.
If you watch the demo video you can see how they would get this: the model is not aggressive enough. While it doesn't cut you off, which is nice, it also always waits an uncanny amount of time to chime in.
The first time I met Tavus, their engineers (incl Brian!) were perfectly willing to sit down and build their own better Infiniband to get more juice out of H100s. There is pretty much nobody working on latency and realtime at the level they are, Sparrow-1 would be an defining achievement for most startups but will just be one of dozens for Tavus :)
> perfectly willing
dreaming
Maybe infiniband is a bit more than we can handle. That technology is incredible! You are right though, we have been willing to build things we needed that didn’t exist yet, or were not fast enough or natural enough. Sparrow-1, Raven-1, and Phoenix-4 are all examples that, and we have more on the way.
Hey @code_brian, would Tavus make the conversational audio model available outside of the PALs and video models? Seems like this could be a great use case for voice-only agents as well.
You can reach out to our sales team. You can chat with our AI SDR here, and they will review it and reach out. https://www.tavus.io/demo
Awesome. We've been using Sparrow-0 in our platform since launch, and I'm excited to move to Sparrow-1 over the next few days. Our training and interview pre-screening products rely heavily on Tavus's AI avatars, and this upgrade (based on the video in your blog post) looks like it addresses some real pain points we've run into. Really nice work.
That’s great! I also built Sparrow-0, and Sparrow-1 was designed to address Sparrow-0’s shortcomings. 1 is a much better model, both in terms of responsiveness and patience.
I am always skeptical of benchmarks that show perfect scores, especially when they come from the company selling the product. It feels like everyone claims to have solved conversational timing these days. I guess we will see if it is actually any good.
Different industry, but our marketing guy once said "You know what this [perfect] metric means? We can never use it in marketing because it's not believable"
Just include some noise, it’s like the most available resource in the universe
Never thought of noise as a resource, but yea.
You should be skeptical, and try it out. I selected 28 long conversations for our evaluation set, all unseen audio. Every turn taking model makes tradeoffs, and I tried to make the best tradeoffs for each model by adjusting and tuning the implementations. I’m certainly not in a position as the creator of Sparrow to be totally objective. However we did use unaltered real conversational audio to evaluate. I tried to find examples that would challenge Sparrow-1 with lots of variation in speaker style across the conversations.
Btw while I think this is cool and useful for real time voice interfaces for the general populace, I wonder if for professional users (eg a dev coding by dictating all day), a simple push to talk is not always going to be superior, because you can make long pauses while you think about something, this would creep out a human, but the AI would wait patiently for your push to talk.
As a dev myself, I see a couple of modes of operation: - push to talk - long form conversation - short form conversation
In both conversational approaches the AI can respond with simple acknowledgements. When prompted by the user the AI could go into longer discussions and explanations.
It might be nice for the AI to quickly confirm it hears me and for it to give me subtle queues that it’s listening: backchannels: “yeah”, and non-verbal: “mhmm”. So I can imagine having a developer assistant that feels more like working with another dev than working with a computer.
That being said, there is room for all modes, all at the same time, and at different times shifting between them. A lot of time I just don’t want to talk at all.
> Non-verbal cues are invisible to text: Transcription-based models discard sighs, throat-clearing, hesitation sounds, and other non-verbal vocalizations that carry critical conversational-flow information. Sparrow-1 hears what ASR ignores.
Could Sparrow instead be used to produce high quality transcription that incorporate non-verbal cues?
Or even, use Sparrow AND another existing transcription/ASR thing to augment the transcription with non-verbal cues
This is a very good idea. We currently have a model in our perception system (Raven-1) that performs this partially. It uses audio to understand tone and augment the transcription we send to the conversational LLM. That seems to have an impact on the conversational style of the replicas output, in a good way. We’re still evaluating that model and will post updates when we have better insights.
It sounds really cool, but I don't see any way of trying the model directly. I don't actually want a "Persona" or "Replica" - I just want to use the sparrow-one model. Is there any way to just make API calls to that model directly?
Such things were doing a good-enough job scamming the elderly as it is--even with the silence-based delays.
That’s unfortunate and certainly not what I spend my time dreaming about. My favorite use case for the elderly is as a sort of companion for sharing their story for future generations. One of our partners uses our technology to help elderly. But yeah, this kind of technology makes AI feel more natural, so we should be aware of that and make sure it’s used for good.
How does it compare with https://github.com/KoljaB/RealtimeVoiceChat , which is absent of the benchmark ?
That's not a turn-taking model, it's just a silence detection Python script based on whatever text comes out of Whisper...
I haven’t tried that one yet, I’ll check it out.
Any examples available? Sounds amazing.
Try out the PALs: they all use Sparrow-1. You can try Charlie on Tavus.io on the homepage in one of the retro retro-styled windows there.
Literally no way to sign up to try. Put my email and password and it puts me into some wait list despite the video saying I could try the model today. That's what makes me mad about these kind of releases is that the marketing and the product don't talk together.
try signing up for the API platform on the site. You can access it there
How do I try the demo for Sparrow-1? What is pricing like?
You can try Sparrow-1 with any of our PALs, or by signing up for a developer account.
I tried talking to Claude today. What a nightmare. It constantly interrupts you. I don’t mind if Claude wants to spend ten seconds thinking about its reply, but at least let ME finish my thought. Without decent turn-taking, the AI seems impolite and it’s just an icky experience. I hope tech like this gets widely distributed soon because there are so many situations in which I would love to talk with a model. If only it worked.
Agreed. English is not my native language. And I do speak it well, it's just that sometimes I need a second to think mid-sentence. None of the live chat models out there handle this well. Claude just starts answering before I've even had the chance to finish a sentence.
English is my native language, and I still have this problem all the time with voice models.
Anthropic doesn't have any realtime multimodal audio models available, they just use STT and TTS models slapped on top of Claude. So they are currently the worst provider if you actually want to use voice communication.
It's unfortunate though, because Anthropic LLMs and ecosystem is the best IMHO. Tavus (we) and Anthropic should form a partnership.
I think Anthropic currently has a slight edge for coding, but this is changing constantly with every new model. For business applications, where tool calling and multi-modality matter a lot, OpenAI is and always has been superior. Only recently Google started to put some small dents in their moat. OpenAI also has the best platform, but less because it is good and more because Google and Anthropic are truly dismal in every regard when it comes to devx. I also feel like Google has accrued an edge in hard-core science, but that is just a personal feeling and I haven't seen any hard data on this yet.
I love Anthropic's models but their realtime voice is absolutely terrible. Every time I use it there is at least once that I curse at it for interrupting me.
My main use case for OpenAI/ChatGPT at this point is realtime voice chats.
OpenAI has done a pretty great job w/ realtime (their realtime API is pretty fantastic out of the box... not perfect, but pretty fantastic and dead simple setup). I can have what feels like a legitimate conversation with AI and it's downright magical feeling.
That said, the output is created by OpenAI models so it's... not my favorite.
I sometimes use ChatGPT realtime to think through/work through a problem/idea, have it create a detailed summary, then upload that summary to Claude to let 4.5 Opus rewrite/audit and come up with a better final output.
I use Claude Code for everything, and I love Anthropic's models. I don't know why, but it wasn't until reading this that I realized: I can use Sparrow-1 with Anthropic's models within CVI. Adding this to my todo list.
Agreed. I tried using Gemini's voice interface in their app. It went like this:
===
ME: "OK, so, I have a question about the economics of medicine. Uh..." [pauses to gather thoughts to ask question]
GEMINI: "Sure! Medical economics is the field of..."
===
And it's aggravated by the fact that all the LLMs love to give you page-long responses before it's your turn to talk again!
Am I not allowed to cut you off if you're ramble-y and incoherent?
Its rude if you're a human, and entirely unacceptable if you are a computer.
The one thing that really surprised me, the thing I learned that's affected my conversational abilities the most: turn taking in conversation is a negotiation: there are no set rules. There are protocols: - bids - holds / stays - implications (semantic / prosodic)
But then the actual flow of the conversation is deeply semantic in the best conversations, and the rules are very much a "dance" or a negotiation between partners.
That's an interesting way to think about it, I like that.
It also implies that being the person who has something to say but is unable to get into the conversation due to following the conversational semantics is akin to going to a dance in your nice clothes but not being able to find a dance partner.
Yeah, I can relate to that. Maybe it's also because you are too shy to ask someone to dance. I think I learned that lesson: just ask, and be unafraid to fail. Things tend to work themselves out. Much of this is experimentation. I think our models need to be open to that: which is one cool thing about Sparrow-1: it's a meta-in-context learner. This means that when it try's and fails, or you try and fail, it learns at runtime to adapt.
Metric | Sparrow-1 Precision 100% Recall 100%
Common ...
The response timing in the chart in the blog post shows that even with perfect precision/recall Sparrow-1 also has the fastest true positive response times.
The turn taking models were evaluated in a controlled environment with no additional cascaded steps: LLM, TTS, Phx. This matters to get apples to apples comparison: without the rest of the pipeline variability influencing the measurements.
The video conversation examples are sparrow-1 within the full pipeline. These responses aren’t as fast as sparrow itself because the LLM, TTS, facial rendering, and network transport also take time. Without Sparrow-1 they would be slower. Sparrow-1 enables the responses being as fast as they are, and with a faster CVI pipeline configuration the responses can be as fast as 430ms in my testing.
If you watch the demo video you can see how they would get this: the model is not aggressive enough. While it doesn't cut you off, which is nice, it also always waits an uncanny amount of time to chime in.
That should lead to a low recall: too many false negatives. I wonder how they are calculating it.
What is "ASR" - automatic speech recognition?
Ah good question: Yes, ASR stands for Automatic Speech Recognition.