Silencio | Voice Data for AI

Level-up your Earnings with $SLC In coordination with the @FoundationSLC, we are unlocking enhanced $SLC utility starting today. SLC holders can now earn up to 100% more through Silencio Voice AI. Live and available to all Silencio Voice AI users as of now. What is it? 🔹Hold more $SLC → Move up levels 🔹Higher level → Higher % boost on your rewards 🔹Up to 100% boost at the highest level Why this matters: 🔹Direct earning power tied to $SLC holdings 🔹Stronger flywheel between enterprise revenue, token burns, and contributor rewards 🔹A clear economic reason to hold $SLC on your wallet Much more is coming to continue driving $SLC utility and placing the token at the absolute core of the Silencio ecosystem. Every new dataset, every enterprise contract, every incentive layer will increasingly reinforce SLC as the coordination engine of the network. We are building the world’s ears for AI and robotics. Visit Voice AI today and start earning: ai.silencio.store/opportunities

Mid-July update on what voice AI builders are most asking us for. No company names, no money figures, just the shape of what's wanted. Top of the list: recordings of free conversation in languages where voice AI struggles, with a checked transcript and the regional variant of the language marked. Next: recordings of more than one person talking together, with clear marks of who said what. Then: recordings where someone switches between languages mid-sentence, with the switch points marked. Then: vocabulary that voice AI usually mangles. Medical terms. Specialist terms. What's NOT being asked for right now: huge piles of unmarked audio. Generic English. Studio-clean recordings. The hard, careful, well-labeled stuff is what's valuable this summer. If you're recording on the Silencio Voice AI Platform, those slow careful recordings matter more than the fast easy ones.

WHAT THE VOICE? Here is what is happening in Voice AI over the past week (and a few days) ranked by industry impact: 1/14 @OpenAI releases GPT-Live-1 and GPT-Live-1 mini — Full-duplex voice models that can listen and speak simultaneously, enabling more natural conversations, interruptions, and live translation in ChatGPT. 2/14 US Senators revive the AI Labeling Act — Bipartisan bill requiring clear disclosures for AI-generated audio, video, and images to improve transparency and combat deepfakes. 3/14 @ElevenLabs explores $22B tender offer — Company is in talks for a secondary share sale that would roughly double its valuation from the previous $11B round. 4/14 @xAI launches Voice Agent Builder — No-code platform to create and deploy production voice agents with telephony and guardrails in under two minutes at $0.05 per minute. 5/14 Cartesia releases Ink-2 — New state-of-the-art streaming Speech-to-Text model optimized for voice agents, leading benchmarks in accuracy and turn detection. 6/14 @GradiumAI raises $100M seed round — Paris-based voice AI startup extends its seed to $100M with Nvidia joining as an investor and plans Bay Area expansion. 7/14 @RetellAI launches Conductor — Graph-native AI copilot that helps teams build, test, review, and improve production voice agents using natural language. 8/14 @DecagonAI launches Decagon Voice — Enterprise-grade AI voice agents for phone support, powered by the same platform as its chat agents and using ElevenLabs voices. 9/14 @Heypocket raises $11M — YC-backed startup behind a $129 AI note-taking hardware puck ships over 130,000 units and reaches strong revenue traction. 10/14 Lucida AI raises €6.1M seed — Speech-to-speech language coaching platform reaches 3 million users and secures funding to expand globally. 11/14 @SynthflowAI AI deploys voice agents at Nellis Auction — Successfully automates 80% of customer interactions with a one-day deployment. 12/14 @Syntiantcorp and Vibe.co partner — Collaboration to bring voice-enabled AI to smart workspace hardware using edge AI chips. 13/14 @healthlynked launches AI healthcare platform — Introduces 24/7 scheduling and medical office voice agents. 14/14 @Forbes analysis highlights patient trust as the biggest barrier — Report argues that trust, rather than technology, remains the main obstacle to healthcare voice AI adoption.

A small reminder for anyone recording on the Silencio Voice AI Platform. The single most important detail to get right on your recording is what language you spoke in. If that's wrong, every other recording grouped with yours becomes harder to use. The AI that learns from it gets confused. Next most important: your regional way of speaking, if your language has one. Then the country you grew up in. Then the kind of room you were in. Then your age range and gender. (Your device fills in automatically.) Each one takes twenty seconds of attention from you. And each one makes weeks of cleaner training possible for the AI later. Twenty seconds, real impact. The most useful thing you can do every time you record.

@messerer_theo 👀👀

Last week we closed the biggest contract in Silencio's history with the #1 Voice AI lab in the market. This is the strongest product-market-fit validation we have received to date. Excited for what we as a team have accomplished and what's ahead for @silencioNetwork. We have received an incredible order for 5000 hours of single & multi speaker Tagalog/Cebuano (Philippines). More languages will follow once we deepen the relationship. Elite Tagalog/Cebuano will receive an email about it. More news to follow soon.🫡

"We support Spanish" or "we support Arabic" is a marketing claim, not a deployment fact. The real question is which regional way of speaking. Recordings broken out by region are what close the gap. That's what we collect for. 3/3

Spanish in Mexico, Argentina, Spain, Colombia, the United States and the Andes are different enough that a voice AI trained on one mostly fails on another. Same for Arabic across MSA, Egyptian, Levantine, Maghrebi, Gulf. Same for Hausa Eastern vs Western. Same for Hindi vs Hindi mixed with English. 2/3

Most languages are not one thing. Treating them as one is one of the biggest reasons voice AI keeps failing. 1/3

Two kinds of recording sit at the core of how voice AI gets trained, and they do very different things. Scripted recordings, where you read a known sentence aloud, give the model a perfect audio-text pairing. Easy to use, fast to train on, and the only kind of data that works for building text-to-speech. Free speech, where you talk about whatever you want, gives the model the way people actually talk. The hesitations, the false starts, the conversational pace, the slang. The only kind of data that works for building voice products that will be deployed against real users. Production-grade voice AI needs both. The Silencio catalog ships both. If you've ever wondered why both formats exist on the Silencio Voice AI Platform, that's why.

For languages that voice AI usually does badly on, the problem isn't that nobody speaks them. People do, plenty. The problem is the whole package around the audio. Transcripts that nobody wrote. Details that nobody captured. Consent nobody documented. The Silencio Voice AI Platform collects the whole package. That's why we exist. 3/3

What AI builders actually need: the audio, a transcript (or label) of what was said, the details about it (language, country, your regional way of speaking, the device, the environment, age range, gender), and clear consent that the recording can be used. Missing any one, and the hour is much less useful. 2/3

What does an hour of voice recording actually need to be useful to AI builders? The audio by itself isn't enough. There's a whole package around it. Worth unpacking. 1/3

The final 4 words are in. (Same Procedure as every month) Each vote count is turned into a number and combined into one string and hashed with SHA-256 to generate a provably random seed. That seed selects the winners. Vote now ⬇️

A note on what voice AI companies are most asking us for this summer. A year ago they mostly wanted more hours of audio. That's no longer the hard part. What's hard now is hours of audio where someone has gone through and checked the transcript, in languages where voice AI still gets things wrong. The second big ask is recordings broken out by region. "Spanish" isn't enough anymore. They need Mexican Spanish, Argentinian Spanish, Spain Spanish, separately. Same for Arabic, same for Hindi, same for many others. The third is recordings of more than one person talking together, with clear marks of who said what. The Silencio Voice AI Platform is being directed against exactly those three. Hours and hours of new voice audio enter our catalog every day. What kind of hours changes based on what voice AI builders are telling us they need.

@coinbureau 1/200

🇪🇺 EUROPE'S CRYPTO LICENSING DEADLINE HITS WEDNESDAY From July 1, unlicensed platforms must stop serving EU users. Any platform without a CASP licence must cease EU operations or risk fines and prosecution. Just over 200 crypto firms including Coinbase, Kraken, OKX and Crypto .com have managed to secure a license.

As the MiCA transitional period comes to an end, many projects are now focusing on regulatory compliance. Together with the @FoundationSLC, transparency and regulatory clarity have been a priority from the beginning. Our MiCA crypto-asset whitepaper has already been published and is available: whitepaper.silencio.network We're focused on building the World's Ears for AI & Robotics.

Why do voice AI systems trained on 100 voices fail in the real world, while ones trained on millions of voices work everywhere? Because a voice carries dozens of things the AI has to learn to ignore. Pitch. Tone. Accent. How fast you talk. Where in the mouth you make a sound. Whether you're excited, tired, congested. Whether you're 8 or 80. Whether you grew up speaking the language or learned it later. If the AI only ever hears one slice of that, say 30-year-old American men reading in a studio, it learns "this is what speech sounds like" and then breaks on anyone outside that slice. This is not theoretical. It's exactly why older voice assistants were notoriously bad at women's voices, children's voices, regional accents, and elderly speakers. The fix is boring but it works. The training has to look like the people who'll actually use the AI. 2 million contributors across 180+ countries gets you closer to that than anything else out there. Every contributor adds one more direction the AI can no longer be fragile in.

Fixing this needs recordings of people actually mixing languages mid-sentence, with the switch points marked. That kind of recording basically does not exist at scale. If you mix languages in your normal speech, recording yourself just talking the way you actually talk is incredibly valuable. 3/3

Most voice AI is trained on one language at a time. It has to pick which language it's hearing before it can do anything. When you switch mid-sentence, it either freezes, refuses, or makes up words. Hindi-English, Tagalog-English, Swahili-English, Spanglish, Arabic-French. None of these work well today. 2/3

Many of you switch between languages mid-sentence without even thinking about it. "I went to the market and bought iyán." "Vou para casa, parking is full." Totally natural way to talk. And one of the hardest things for voice AI to handle. 1/3

Roughly 40% of the world's languages are considered endangered. Many won't have a fluent speaker by the end of this century without intervention. Recording them, in real conditions and with proper metadata, is one of the most direct things technology can do for them. We're trying not to lose any of them. 2/2

A language is not just a way of communicating. It carries a way of thinking, a way of joking, a way of arguing, a way of grieving. There are concepts in some languages that have no clean equivalent in any other. If a language disappears, those concepts don't move to another language. They are simply gone. 1/2