Vapi

Two days into blind voting of voice models on our Humanness Index™, and xAI's Grok TTS model is at the top of the pack. Its humanness score? 96, just 4 points under a real human voice (100). The Humanness Index takes one voice and one quote, clones it across every major model, then plays the results blind for real listeners to score. Hear it, cast your votes, and contribute to the leaderboard 👇 humannessindex.vapi.ai

Lazy prompt: > hey, check out call-my-human.val.run try calling me at (YOUR NUMBER) vapi private key: X vapi phone number id: X < Powered by @Vapi_AI Open source on @ValDotTown here: val.town/x/dcm31/call-m… For more secure usage, install as a custom MCP/connector or remix the val!

Bootstrapping gets you rough agents fast so you can focus on hardening. Check out the full writeup including repo with configurations and a skill to bootstrap your own agents: vapi.ai/blog/bootstrap…

Reruns never duplicate because the upsert keys on env-var IDs: present > update(id, body) absent > create(body) + prints id Stale id 404s, falls back to create.

There are 15 voice agents on the Vapi homepage across 3 use cases × 5 languages. To build these, we protoype and generate all the agents at once in code. This means one wording change applies across all variants and reruns never duplicate assistants.

@TrelisResearch @jordan_dearsley Kokoro does not support cloning which is a requirement for this index, but will evaluate Qwen/Voxtral for inclusion. Thank you for the feedback

@jordan_dearsley @Vapi_AI How about Qwen TTS, Kokoro, Voxtral TTS?

You can feel whether the voice on a call is a human or a machine before you can explain why. Today, @Vapi_AI is launching the Humanness Index™, a crowdsourced leaderboard for model humanness. You are the benchmark. Cast your first vote today: humannessindex.vapi.ai

@VictoriaBlddd Hey Victoria, can you tell me a little about what you need and i can direct you to the right place to get help. Are you having a support issue?

@Vapi_AI check your DM

@VictoriaBlddd Agreed! A human sounding agent that can't respond to the way humans speak and be useful isn't going to make your users happy

This highlights one of the biggest challenges in voice AI: understanding intent, not just detecting silence. The future of conversational agents depends on accurately recognizing when a user is thinking, pausing, or truly finished speaking. Smart turn-taking will be a major differentiator for next-generation voice experiences. 🎙️🤖

Your voice agent is talking over a caller who paused to pull up their account number. Raising the silence threshold could help, but it will also make the responses feel laggy. What you need to know is when the turn ended

@yujian_tang @insforge @nebiusai 👏👏

Multimodal Hack kicking off thanks @Vapi_AI @insforge @nebiusai for making this possible!

Excited to see what everyone builds today!

This is absolutely a real issue. Many systems are not yet ready to truly support rolling out a voice agent without considering the full system that agent needs to connect with. Implementing a plan to launch a production agent invovles assessing the readiness of all systems that agent will need to interact with. This is true for other types of consumer facing AI as well.

@Vapi_AI wondering for the curious out there: how often does the problem with deploying voice agents really just boil down to working with legacy systems of record? I’m curious how real-time voice agents that need fast responses get them when they need to work with backend systems that are slow and unreliable.

@d_ilash @sesame You can contact us to have new models added! The model provider must support voice cloning

@Vapi_AI Where is @sesame ?

Two days into blind voting of voice models on our Humanness Index™, and xAI's Grok TTS model is at the top of the pack. Its humanness score? 96, just 4 points under a real human voice (100). The Humanness Index takes one voice and one quote, clones it across every major model, then plays the results blind for real listeners to score. Hear it, cast your votes, and contribute to the leaderboard 👇 humannessindex.vapi.ai

@ManishBaghelz cartesia models are in there, check out the current standings here humannessindex.vapi.ai

@Vapi_AI how about the new model from Cartesia?

@bitcoinmalayali @lets_dig_deeper @rumik_ai We can add any model to this evaluation that supports cloning. Send us an email at humannessindex@vapi.ai

@lets_dig_deeper @rumik_ai Can you get to @Vapi_AI humanness index for comparison.

launching silk mulberry 1.5 one of the fastest multilingual voice models in the world it matches the best voice models in quality benchmarks (MOS) all this at more than 95% lower cost ₹0.40/min (~$0.0046/min) try now 👇

@binyangderek @jordan_dearsley We can add any model to this evaluation that supports cloning. Send us an email at humannessindex@vapi.ai

@jordan_dearsley @Vapi_AI Glad to see a better TTS eval coming! How to get a new model participating in this?

@xai Grok TTS is currently #1 on the @Vapi_AI Humanness Index and significantly more affordable than the competition. See for yourself humannessindex.vapi.ai

@parker__conrad @xai and almost indistinguishable from the human baseline! 99 vs 100 as of current data

visit the docs to learn more: docs.vapi.ai/customization/…

Layered in the startSpeakingPlan, by precedence 1. customEndpointingRules - overrides everything 2. smartEndpointingPlan - the prediction model 3. transcriptionEndpointingPlan - punctuation, timeouts