Apoorv Sood

This Mumbai cafe manager has the most unexpected story At 2:30 PM, three guests came for breakfast and kept waiting outside the cafe. The staff had no idea who they were waiting for. Around 6 PM, Christopher Nolan, Tom Holland and Matt Damon finally arrived. They had chai, bun maska and some light snacks. Within minutes, police, bodyguards and a huge crowd showed up, so they left after just 10-15 minutes. The best part? The manager only Googled them later and said, "Tab pata chala Tom toh Spider-Man wale actor hain." 😭☕ Imagine serving Christopher Nolan, Spider-Man and Jason Bourne... and finding out after they leave! #TheOdyssey

Pulse STT is now live as a routable speech-to-text provider on the Speko (dot) ai gateway! One config change is all it takes to route your voice agents to Pulse: no separate integration or extra API keys required. If you're building agents where responsiveness is the product, Pulse STT is the first model Speko would point you at. @beknabdik @spekoai_

Lightning 3.1 Pro from @smallest_AI is the #1 real-time TTS model for Customer Support in Hindi 🇮🇳 on Voice Arena In blind human evaluations it outperformed leading Hindi alternatives including Sarvam and Cartesia. ➤ Quality: Bradley-Terry Elo of 1101 (±24) from 478 blind arena appearances in this category, every point earned in head-to-head preference tests, not internal benchmarks or synthetic metrics ➤ Production-ready: native 44.1 kHz output, ~150 ms time-to-first-byte at scale, ~29 languages with premium curated voices, purpose-built for voice agents, IVRs and support workloads Customer support is the register where TTS actually gets deployed. Samples below 🧵

Thrilled to have been part of the ELETS NBFC Tech Summit in Chennai! 🎙️ It was incredible to be in the room and have the opportunity to showcase realtime Voice AI deployments. A few takeaways that stuck with us: → NBFCs are moving fast on AI adoption, but the real differentiator is execution → Trust and compliance aren't blockers. They're the foundation → The appetite for intelligent, voice-first customer experiences is real and growing To everyone we connected with, thank you for the energy, the ideas, and the honest conversations. The future of fintech in India is being written right now, and we're glad to be part of it.

Incredible energy at the Elets Banking and Finance NBFC Tech Summit in Chennai! Our Head of Global GTM, @apoorvsood, took the stage to share how voice AI is already transforming collections, onboarding, and customer engagement for NBFCs. The conversations in this space are getting serious, and we're here for it. If you want to see what a voice AI deployment looks like for your NBFC, let's talk. Book-a-demo link in the thread

We're live at the ELETS NBFC Tech Summit in Chennai! Come find us at Booth 16. We're running live demos all day and would love to show you what's possible with voice AI in financial services. Later today, our Head of Global GTM @apoorvsood takes the stage to walk through a real voice AI deployment built specifically for NBFCs: the architecture, the results, and what's next. If you're here, don't miss it. If you're not, find the book-a-demo link in the thread. 📍 Hilton Chennai | Booth 16

Pulse STT realtime is ahead of @DeepgramAI @cartesia and @ElevenLabs on speko benchmarks! The world’s best STT to build real-time voice agents!

#1 Speech-to-Text model in the world on Speko Benchmarks! With an industry-leading 183ms end-of-turn latency, Pulse tops the leaderboard with the fastest conversational experience, delivering responses 2× faster than the next-best model while maintaining super strong transcription accuracy. And starting today, Pulse STT is also live as a routable provider on the @spekoai_ Gateway. If you're building real-time voice agents where responsiveness is not a metric but the product itself, Pulse is the model Speko recommends! Read Speko's deep dive on Pulse STT in the thread below. @beknabdik

This week in conversational AI: Here is essential weekend reading if you build voice: ElevenLabs — Procedures in ElevenAgents: SOP-style playbooks that load only when triggered to steer agent behavior. An agent-control layer, not just TTS. Gradium — Translate: Two real-time translation models, one transcribes and translates in a single pass, the other translates speech-to-speech while matching the user's voice. Cartesia — channel-aware voice preview: Hear how a voice will sound on telephony vs the web before you ship it. xAI — Grok Voice Agent Builder (beta): No-code voice agents on Grok's native speech-to-speech path, with telephony, guardrails, and observability built in, at $0.05/min. A frontier lab just entered the platform market at a price floor and its voice, TTS, and STT models also showed up in Vercel's AI Gateway, so you can call them without juggling separate xAI keys. Retell AI — Conductor: A copilot that reviews your calls and writes simulation tests; it's already authored 70% of Retell's own. Point it at a call that went wrong and it turns that into a test case, suggests a fix, and re-runs the sims before anything ships. Soniox — Voice Cloning: Production-grade voice replicas in 60+ languages from just a few seconds of audio. Hugging Face × Cerebras — open speech-to-speech pipeline: A fully open, swappable stack on Gemma 4 that tackles tail latency (P95), not just the average, which is what actually makes a call feel snappy. NVIDIA × Palantir are bringing Nemotron open models into air-gapped government and critical-infrastructure environments. India's VoicERA launched a government-backed, open-source end-to-end voice stack on BHASHINI. smallest.ai Pulse cites and tops Coval's STT leaderboard at 2.7% WER. Providers are increasingly proving relevance on an independent board anyone can reproduce. Rime × Miravoice published a ~100k-call field study on which AI voice actually keeps people on the phone: 12 voices across 3 providers, scored on completions and hangups. Pipecat refreshed its classic Voice AI & Voice Agents primer. The headline: turn detection is basically solved via hybrid VAD + Pipecat's Smart Turn model + LLM single-token tagging, and open-weight LLMs are redrawing the speed/quality curve. Standing up a voice agent is becoming nearly no-code: xAI's Grok Voice Agent Builder and Retell's Conductor both ship describe-it-in-plain-English creation, with ElevenLabs Procedures bringing that natural-language approach one layer in, steering how an existing agent behaves. As building commoditizes, the real differentiator moves to testing and trust.

Happy 4 of July! Let the fireworks begin…

The future of the firm is a learning loop in which human capital and token capital compound. With our new Frontier Co., our ambition is to help every enterprise build its own AI capability, and to help create a frontier ecosystem where every organization can turn its knowledge, workflows, and judgment into its own AI systems that continuously improve. blogs.microsoft.com/blog/2026/07/0…

Which one was actually up there? @smallest_AI

The voice agent can listen to what you said. But not how you said it. That's why we first looked into speech-to-speech. But the deeper reason being on how the brain actually works. It thinks while listening, not after. Got into where this goes next on the fireside chat with @realbasilchatha

Small Models can be smarter. We didn't buy that scaling laws alone get you to human intelligence. And now we’re researching on how to get asynchronous models to mimic the human brain Full chat with @realbasilchatha linked below👇

Last week, the US team flew to Vegas for CCW, the biggest conference on our calendar this year. Six people on the ground. Months of prep. Two days on the floor. The pitch was simple: Smallest AI builds the fastest, most human-sounding voice AI in the market, and unlike most of what you see, we own the entire stack. Our own speech-to-text, our own language model built specifically for voice, our own text-to-speech. No third-party APIs stitched together. One team accountable for the whole system, and nothing leaves the customer's environment.For banks and enterprises running millions of phone calls - support, collections, sales, scheduling, that distinction matters. We demoed Avery live on the floor – our voice agent handling the calls these teams field every day. Healthcare intake. Insurance. Retail order inquiries. Banking and card servicing. Conversations specific to each industry. If you've deployed voice AI before, you already know the failure modes. It talks over your customers. It stalls the moment a call goes sideways. It sounds like a machine exactly when the conversation needs a human. At low volume you can paper over it. At enterprise scale, there is no room for a system that buckles under load. That's the problem we built the stack to solve. We also took a bet on a side event. Together with @telnyx , we hosted a steakhouse dinner with 60 senior leaders in the room. Every invite was deliberate. Conversations that started during the day continued over dinner, and new faces from that table came back to learn more the next morning. Everyone left with a leather bag engraved with Smallest AI. None of it happens without the six who ran it on the ground, fully, for two days. By the end of day two, we'd already decided: we're coming back to CCW. And we're not done. Half the team is now headed to NENA 911 in Columbus, June 29 to July 2, to talk transcription accuracy on real 911 audio at scale. If you run a PSAP or build NG9-1-1, call-handling, logging/QA, or CAD infrastructure, find Harsh Jain and Alex Ducamp on the ground. @apoorvsood @lilyyang169

Voice AI has a blind spot. It hears your words, but not your tone. Feed a model what you said, and it responds. But it can't read the frustration, the hesitation, the sarcasm. "I'm fine" said calmly and "I'm fine" said through gritted teeth land exactly the same. That's the biggest problem facing voice AI today. See how we're solving that in the link below.

Pulse just hit 2.8% WER on @covaldev STT leaderboard. What makes this different from every other benchmark: Coval tests streaming, not batch. Real-time WebSocket audio, simulated at actual call pace - the same path your voice agent runs in production. For BFSI, that gap matters. A model that aces offline benchmarks and chokes on a live call is a demo, not a product. Pulse was built for the call, not the lab. The benchmark just proved it.

Today, Valence is proud to announce that we have raised a $5M seed round to build emotional intelligence infrastructure for voice AI.

Day 1 of CCW brought exactly the kind of conversations we came for. The floor was packed with senior leaders across industries, all curious to learn more about what's new in voice and AI. All day at booth 1620, we demoed Avery, the @smallest_AI voice agent live across four use cases: healthcare clinic intake, insurance agent, retail order inquiry, and banking card servicing. After the conference, we also co-hosted the Steakhouse After Hours with @telnyx at Brasserie B. It was refreshing to see everyone chatting comfortably with each other over good food and drinks. As a gift, everyone also received a branded leather bag. Today is Day 2, the final day of CCW. We'll be at booth 1620 until the floor closes. Lets go team! @apoorvsood @lilyyang169

The @smallest_AI candle tower 🙌 Come visit us at booth #1620 at CCW!