What is Raven-1?
Raven-1 is the perception layer of Tavus's Conversational Video Interface, and it is on by default in every Tavus conversation. It reads the user's tone, pauses, facial expression, gaze and posture, plus any shared screen, and passes short notes such as "the user sounded sarcastic" to the language model with each turn. It replaced the vision-only Raven-0 and is the perception model Tavus recommends.
Where most emotion systems pick one of a few labels, Raven-1 writes free text and can fire your own tools when it spots a cue you describe. It is not sold on its own: you get it with any Tavus plan at no extra charge.
What changed from Raven-0?
Raven-1 adds hearing: Raven-0 only watched the video, while Raven-1 reads voice and video together.
- Audio
- It reads tone and emotion from the user's voice, which Raven-0 could not do.
- Sound and picture together
- Voice and video are read as one, so a smile with a sarcastic tone is read differently from a warm one.
- Setup
- Visual settings carry over; some field names changed, and the old names still work.
- Default
- Raven-1 is now the default perception model for every Tavus PAL.
Who is Raven-1 best for?
Raven-1 is best for teams building Tavus video agents for healthcare intake, coaching, interviews or sales, where the agent has to notice tone, hesitation and facial cues as well as words.
Healthcare intake and therapy
It picks up when a patient sounds anxious or seems to hold something back, so an intake assistant can change its approach or flag the case for a person.
Coaching and training
It tracks frustration or confusion building over a session, so a tutor agent can step in with a hint before the learner gives up.
Interviews and sales calls
You define events such as hesitation on a topic or a buying signal, and it sends a tool call when it spots one.
Developers already on Tavus
It runs by default in every Tavus conversation and describes the user's state in plain sentences your LLM can read directly.
What's included with Raven-1?
Plain-language notes on tone and expression, custom watch-for questions, perception tool calls, screen reading, an end-of-call analysis and an EU emotion switch.
- Plain-language notes
- Each user turn gets a short note on tone and emotion, up to 32 tokens, sent to the LLM with the transcript.
- Custom watch-for questions
- You list things to keep checking in the video or audio, such as whether the user looks distressed, and it checks the video about once a second.
- Perception tool calls
- Vision and audio tools fire when it sees or hears a cue you describe, running in parallel so the agent keeps talking.
- Screen reading
- When a user shares a screen in a video call, it reads that screen the same way it reads the camera, with no extra setup.
- End-of-call analysis
- Questions you set, such as how often the user looked at the screen, are answered once the call ends and sent to your callback.
- EU emotion switch
- An emotion_recognition setting stops emotion reading from faces and voices, and it switches off by default for conversations marked with the EU policy.
How does Raven-1 work?
Start a Tavus conversation
Raven-1 is on by default for every PAL; set perception_model to off only if you do not want it.
Say what to watch for
Add visual and audio awareness questions, or register vision and audio tools that fire on a cue.
Read what it noticed
Notes arrive with each user turn as user_visual_analysis and user_audio_analysis, and the end-of-call analysis goes to your callback.
How much does Raven-1 cost?
Raven-1 comes with every Tavus plan at no extra charge: the Free plan includes 20 minutes of conversational video a month, and paid plans cost $22, $59, $397 or $975 a month.
- 20 minutes of conversational video a month
- Stock faces and 42+ languages
- One conversation at a time, up to 5 minutes each
- 60 minutes of conversational video a month
- 1 custom face, no watermark
- No pay-as-you-go minutes past the allowance
- 175 minutes a month, then $0.35 a minute
- 3 custom faces
- Up to 3 conversations at once, 15 minutes each
- 1,300 minutes a month, then $0.31 a minute
- 7 custom faces and 100+ stock faces
- Up to 10 conversations at once, with recordings
Business costs $975 a month for 4,000 minutes, 15 custom faces and 15 conversations at once, with extra minutes at $0.26 and priority support. Enterprise is priced on request with volume discounts. Raven-1 is part of Tavus's Conversational Video Interface and has no separate price. Each conversation is billed in 6-second steps with a 30-second minimum, and recordings cost $0.03 a minute. Full pricing on tavus.io. We checked these prices on .
How do readers rate Raven-1?
No one has rated Raven-1 yet. Be the first.
Sign in to rate. One rating per account.
Raven-1 pros and cons
It gives Tavus agents a running read of tone and facial cues at no extra cost, but it only runs inside Tavus and its emotion reading is restricted in EU workplaces and schools.
Pros
- Costs nothing extra: it runs in every Tavus plan, including the free 20 minutes a month, and Tavus bills only conversation minutes.
- Writes plain sentences the language model can use directly, instead of a handful of emotion labels, and tracks how a user's mood shifts within one turn.
- Quick enough for live calls by Tavus's figures: audio notes in under 100ms, and a combined read that is never more than 300ms old.
- Hacker News commenters who tried the launch demo were impressed by it, including some who were uneasy about the idea.
Cons
- It only works inside Tavus: there is no separate Raven-1 API, so you take Tavus's avatars, pipeline and per-minute pricing with it. Hacker News users have asked Tavus for the same direct access to its turn-taking model.
- Emotion reading from faces and voices is restricted in EU workplaces and schools, and Tavus turns it off by default for EU-policy conversations, which limits interview and training use there.
- Commenters worry about bias and misuse: a loud speaker read as upset, an off-center camera read as avoiding eye contact, and hiring teams leaning on its judgement.
- Long sessions get costly or cut off: Free and Starter cap each conversation at 5 minutes, and extra minutes cost $0.26 to $0.35 on the plans that allow them.
What are the best Raven-1 alternatives?
Hume AI for voice-only agents, Gemini 3.8 Live for one model that sees and hears, and LiveAvatar by HeyGen for avatars at scale.
Hume AIFreeChoose it when your agent is voice only: its EVI voice interface measures the tune, rhythm and timbre of speech as it listens, and paid plans start at $3 a month.
Gemini 3.8 LiveFreeChoose it when you want one model that hears and sees the user itself, with no rendered face; it has a free tier, and paid audio costs about $0.005 a minute in and $0.018 out.
LiveAvatar by HeyGenFreeChoose it when scale matters more than perception: its $99 Essential plan has no concurrency cap and lets you bring your own LLM and voice stack.
Questions about Raven-1
What is Raven-1?
Raven-1 is Tavus's real-time perception model for video agents: it watches the camera, hears tone of voice and tells the agent's language model, in plain sentences, how the user seems to feel.
Who is Raven-1 for?
Raven-1 is best for teams building Tavus video agents for healthcare intake, coaching, interviews or sales, where the agent has to notice tone, hesitation and facial cues as well as words.
Is Raven-1 free?
Yes, within limits. Raven-1 runs in every Tavus conversation, including the Free plan, which gives 20 minutes of conversational video a month with each conversation capped at 5 minutes. More minutes start with the $22 Starter plan.
Does Raven-1 cost extra on Tavus?
No. Tavus says each conversation already includes its models, Raven-1 for perception, Sparrow-2 for turn-taking and Phoenix-4.5 for rendering, along with the LLM, speech and video streaming, so you pay only for the conversation minutes on your plan.
What are the best Raven-1 alternatives?
The closest alternatives on Recaply are Hume AI, Gemini 3.8 Live and LiveAvatar by HeyGen.
Where can I try Raven-1?
On its official website, tavus.io.
Can I use Raven-1 without Tavus?
No. Raven-1 runs only as the perception layer of Tavus's Conversational Video Interface, switched on per PAL with perception_model. Tavus does not offer it as a separate model API, though you can build on Tavus conversations through its API, Pipecat or LiveKit.
Can Raven-1 see my screen?
Yes. When someone shares a screen in a Tavus video call, Raven-1 reads it the same way it reads the camera, so the agent can see on-screen text or errors. It needs a live video call; audio-only conversations get no screen reading.
Does Raven-1 work for users in the EU?
Yes, with limits. For conversations created with the EU policy, Tavus turns off emotion reading from faces and voices by default and adds an AI disclosure, while other cues such as speech content and objects still pass through. Tavus says checking your own legal obligations is up to you.
Our verdict
Raven-1 is worth keeping on if you build on Tavus, but it is not a reason on its own to choose Tavus.
It gives the agent a running, plain-language read of tone, expression and screen at no extra cost, and by Tavus's figures it is quick enough for live calls.
- Skip it for
- Voice-only agents, teams that want a perception model to call from their own stack, and EU workplace or classroom uses where emotion reading is restricted.
- Check first
- Test it on the free plan with your own watch-for questions and see whether its reads are accurate for your users before you act on them.
MAI-Transcribe-2-Streaming
GPT-Realtime-2
Scribe v2 Realtime
GPT-Realtime-2.1
Sonic 3.6
Eleven Realtime