Tether, the company behind the world's largest stablecoin, has released TranscribePsy, a 254-million-parameter speech recognition model that transcribes conversations directly on smartphones, smart glasses and humanoid robots without sending any audio to cloud servers. The release, announced this week as part of the company's QVAC artificial intelligence project, is the latest sign that Tether is serious about building AI software that runs entirely on personal devices.
The move puts a crypto company in an unexpected position: competing with cloud AI providers on the argument that voice data is too sensitive to leave the device. "The future of AI shouldn't mean giving centralized platforms a front-row seat to our personal lives," Tether CEO Paolo Ardoino said in the announcement. He described TranscribePsy as an example of speech recognition that can operate fully on personal hardware, with no dependency on centralized infrastructure. For more context on this story, see our ongoing artificial intelligence updates.
Benchmark Results: Faster Than Whisper on Phone CPUs
According to crypto.news, which reviewed the company's published evaluation, Tether tested the model across six smartphones and found it ran 2.6 times faster than the next fastest system evaluated, and 5.7 times faster than OpenAI's open-source Whisper-small model when running on phone CPUs. On the iPhone, a Core ML-optimized version required less than half the disk space and 1.7 times less memory than a competing Core ML model included in the comparison.
Accuracy was measured using word error rate, the standard metric that counts how many words a transcription system gets wrong against the original spoken content. Tether published the evaluation configurations alongside the model so that independent researchers can reproduce the reported numbers — a transparency practice that remains uncommon among consumer AI vendors.
Four Versions for Different Hardware
TranscribePsy ships in four builds aimed at different classes of hardware:
- A laptop model supporting 25 European languages
- A smaller configuration focused on eight commonly used languages
- A GGUF version that runs through the parakeet.cpp engine inside Tether's qvac-fabric-speech software, for mobile devices
- A Core ML version optimized for Apple's Neural Engine, letting compatible iPhones transcribe using dedicated silicon
The models are designed to handle continuous audio, producing text shortly after speech is detected. Tether says the system could support voice-controlled applications that respond without ever transmitting recordings to external servers — a meaningful claim for smart glasses and wearable devices, where always-on microphones have raised persistent privacy concerns.
Part of a Broader QVAC Push
TranscribePsy is the latest release under Tether's QVAC project, which has expanded steadily through 2026. In April, the company introduced the QVAC SDK, an open-source development toolkit for running AI applications on iOS, Android, Windows, macOS and Linux, covering text generation, speech processing, image recognition and translation. The toolkit is built on QVAC Fabric, a customized version of llama.cpp, and uses Holepunch technology for peer-to-peer model distribution.
Earlier in the year, Tether integrated BitNet LoRA training into QVAC Fabric, saying it had fine-tuned models of up to 3.8 billion parameters on consumer devices including the Pixel 9, Galaxy S25 and iPhone 16, with tests covering models up to 13 billion parameters on the iPhone 16. In September, Ardoino disclosed work on a peer-to-peer search engine that has stored the full Wikipedia archive across 100 distributed nodes. And in October 2025, the company released the QVAC Genesis I dataset, 41 billion synthetic tokens aimed at training models in mathematics, science and technical reasoning.
Why It Matters
The on-device AI race has intensified as device makers seek features that work without connectivity and without data leaving the handset. Whisper remains the most widely deployed open speech model, but it was not designed for the power envelope of a phone CPU, which is where purpose-built small models have room to differentiate. Tether's benchmark claims, if they hold up to independent replication, would put TranscribePsy at the front of that pack.
For Tether, the strategy is also a hedge. The company's core business remains the USDT stablecoin, but its leadership has spent 2026 building a portfolio of local-first AI infrastructure — models, SDKs, datasets and peer-to-peer distribution — that aligns with the decentralized ethos of its user base while reducing dependence on the small number of cloud providers that currently control most AI inference.
Whether a stablecoin issuer can become a credible player in speech AI is an open question. But with a published benchmark, reproducible configurations and a working model that runs on the phones people already own, TranscribePsy is a more concrete release than most on-device AI announcements — and a sign that the privacy-first segment of the AI market is attracting unlikely entrants.
The Trade-Off: Capability Versus Control
Small on-device models still trail their cloud counterparts on raw accuracy, particularly on accented speech, background noise and specialized vocabulary. What they offer in exchange is structural: no network dependency, no per-transcription API cost, and no third party holding the audio. For wearables like smart glasses, where a cloud round-trip adds latency to every interaction and an always-listening microphone is the entire point of the product, local processing is arguably the only design that scales without a privacy backlash.
Tether is betting that this trade-off tilts further toward local processing with every hardware generation. Neural engines in current flagship phones already deliver double-digit-billion-parameter inference, and models in the few-hundred-million range — like TranscribePsy — leave headroom for longer context, more languages, and always-on operation without draining the battery.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
