Harsh Pare

building @tazolabs | all things AI | ex-deshaw, @bridgetownai, @EverpureData | @iitdelhi

[Harsh Pare @harsh_pare]

Kalshi is paying someone $120k to play GTA VI

They literally had to mention "This is a real job" on the application page

[Tazo Labs @TazoLabs]

Tool calls can start before the user stops speaking.

This latest NTU paper validates that a speculative call can launch as soon as the audio grounds the tool and all required arguments, without waiting for the end of the turn.

Try it today in the voice AI pipeline with Tazo.

Paper link in the first reply.

[Tazo Labs @TazoLabs]

My foundations are strong, challenge me 💪

[Sachin Kumar Singh @sachin_singh092]

Though benchmarks are definitely not the end goal for a product, they do help in validating the trajectory. And given these numbers, we feel are on the right one.

Our aim at @TazoLabs is to squeeze maximum output from our models and give the most accurate results instantaneously.

And we are just getting started.

[Harsh Pare @harsh_pare]

Excited to launch DECIDE: The first ever benchmark for audio decision models made from production calls data.
Submit your model's API endpoints and evaluate your model on the benchmark.

Thanks to the folks at @bolna_dev for the collaboration.

[Naman Jain @namanjn1998]

Introducing DECIDE: a benchmark for audio decisions in voice agents, in association with @bolna_dev

A voice agent doesn’t just transcribe. It decides: speak or wait, route the call, pick a tool, switch language. Text can’t tell you if the caller sounded annoyed, interrupted, or talked over the agent.

DECIDE tests exactly that, on real production calls.

→ 4,867 decision questions from 1,081 real agent-caller clips (3 to 25 seconds)
→ 32 tasks across 6 families: turn-taking, graph routing, tool decisions, language switching, caller emotion, business intent

Direct-audio models and STT + text models (20 in total) are scored on the same questions, so you can see what audio actually adds, and what it costs in dollars and milliseconds.

Submit your model endpoint and get scored; no invite needed.

Website: https://t.co/FQnnwZXVXb
Samples: https://t.co/M9sIc8V0en

[Maitreya Wagh @maitreya_wagh]

Excited to work closely with the hottest new entry in the voice ai space! @harsh_pare @namanjn1998

[Harsh Pare @harsh_pare]

@namanjn1998 @bolna_dev 🚀🚀🚀

[Tarush Agarwal @tarush_agarwal_]

Introducing Cekura Bench: the world’s leading Voice AI Benchmarks

We benchmarked 50+ models from openAi, Google, Nvidia, xAI, Inworld, Gradium, Assembly, Fish audio and more

Now you can:
→ Compare naturalness, performance and cost
→ See how the best voice agents are being built.
→ Find the best fit for what you are building.

Our tests cover Speech-to-Speech models, Speech-to-Text models, Text-to-Speech Models, and Voice Agent platforms. We recorded real humans doing tasks in different domains and created the most representative voice benchmark out there!

Explore the benchmarks: https://t.co/eZAFy7P0Nj

We would love to collaborate with more teams on these benchmarks. Reach out if you are interested.

[Axiom @axiommathai]

Most think vibecoding at scale is slop. Here’s a counterexample.

Today we release a 16M-line formalization of the entire Stacks Project, critical infra underlying modern math, with optimized code quality and its highly interconnected structure preserved.

A mathematical city 🌆

[Tazo Labs @TazoLabs]

Today marks the day after which every voice agent will start listening.

[Sahit PJ @sahitpj]

Voice in, decisions out - that perfectly describes our vision @TazoLabs . Voice understanding has never been about the words being said, it’s more than that - it includes acoustics, the hmm’s and buts and much more. I believe that having tazo in your voice stack will be a very important piece to unlock additional capabilities in voice agents.

🚀🚀

[Sachin Kumar Singh @sachin_singh092]

We worked really hard to improve model latency and accuracy over the past few weeks. I can't wait to see all kinds of projects on top of @TazoLabs.

Hope you all like it.
Give it a try: https://t.co/w59v0N5286

[Harsh Pare @harsh_pare]

Live calls move quickly. Your voice agent has to make decisions while the conversation is still happening.
We built Tazo to make decisions directly from the audio signals. Tazo takes streaming audio, your prompt, and your options, and returns calibrated probabilities.

Excited to put Tazo (@TazoLabs ) in developers’ hands. Bring your toughest voice-agent use cases.
Give it a try: https://t.co/DozrB1g3Te

[Naman Jain @namanjn1998]

Your voice agent can get every word right and still get the call wrong.

Introducing Tazo (@TazoLabs), an audio decision model for live voice agents. Speech in, decision out.

Give it raw call audio, a text prompt and your options. It returns the calibrated probability for each option, directly from audio.

• 70.9% vs 46.4% for STT + Jev on real production calls
• Ranks #2 on audio intelligence benchmark (MMAU)
• <200 ms model latency
• First decision model tuned to run on streaming voice calls
• Introducing “None of the above” option to handle out-of-scope and unsure states.

API is live. Free credits for every new sign-up.
Give your agent ears @ https://t.co/Czpr63FjlI

See it in action below 👇

[Harsh Pare @harsh_pare]

@namanjn1998 @TazoLabs 🚀🚀🚀

[Harsh Pare @harsh_pare]

Should've taken some notes while watching Silicon Valley

[michael @mchlhess]

should I be offended by this

[Harsh Pare @harsh_pare]

@namanjn1998 LFG 🚀🚀

[Harsh Pare @harsh_pare]

Voice is not simply text with an ASR step in front of it. Speech contains emotions, timing, hesitation, interruption, and turn-taking signals that a final transcript flattens away.
A decision model that can use the live audio signals can make voice agents much more responsive and better timed.

[Naman Jain @namanjn1998]

I think the biggest gap in voice AI is that agents make decisions from a transcript, not from the audio. I wrote down my thoughts.

I’d love to hear where you disagree. https://t.co/hHmfjlbFKG