Have you ever stopped mid-conversation and wondered where those voice recordings actually end up? Most of us talk to our devices every day without giving it a second thought. Yet the idea of private chats, medical notes or confidential interviews sitting on some distant server feels increasingly uncomfortable. That quiet unease is exactly what makes the latest move from Tether stand out.
Why Offline Speech Recognition Suddenly Matters
The company behind the world’s largest stablecoin has released an open-source speech model that works entirely on the device. No internet required. No data leaving your phone, glasses or robot. Called QVAC TranscribePsy, the 254-million-parameter system is designed for exactly the kind of everyday hardware most of us already own or will soon own.
I have to admit, the timing feels right. Smart glasses are slowly moving from sci-fi curiosity to something you might actually wear. Humanoid robots are no longer pure research projects. And people are growing more protective of their spoken words. In that climate, a model that keeps everything local is more than a technical flex. It is a statement about control.
What TranscribePsy Actually Delivers
At its core, the model listens and turns speech into text without ever calling home. On an iPhone 17 Pro Max it handled ten minutes of audio in just 1.65 seconds using Apple’s Core ML framework. That kind of speed changes how you think about voice interfaces. Suddenly real-time transcription on the device itself becomes practical rather than theoretical.
The system comes in several flavors. One laptop version covers 25 European languages. A lighter version sticks to eight widely spoken languages and fits more comfortably on phones. There is also a GGUF build that runs through the parakeet.cpp engine and a dedicated Core ML package tuned for Apple’s Neural Engine. Developers can pick the right size for the hardware they target.
Performance numbers released by the team look strong for a model of this size. It posted an 8.02 percent word error rate on the Open ASR Leaderboard and 6.58 percent on Multilingual LibriSpeech. On Common Voice the score sat at 21.56 percent. More interesting still, it beat several larger models on the Multilingual LibriSpeech test even though those competitors carried three or four times the parameters.
The future of AI shouldn’t mean giving centralized platforms a front-row seat to our personal lives.
– Paolo Ardoino, Tether CEO
That single sentence captures the philosophy behind the whole QVAC project. Local first. Privacy by design. Control stays with the user.
Speed Tests That Raise Eyebrows
Across six different smartphones the model ran 2.6 times faster than the next quickest system in the comparison. It was 5.7 times faster than Whisper-small when both ran on phone CPUs. The iPhone-optimized version needed less than half the disk space of a competing Core ML model and used 1.7 times less memory. Those are not small differences when you are trying to squeeze useful AI onto everyday devices.
I keep coming back to the practical side. If a pair of smart glasses can transcribe a conversation the moment it happens, without lag and without uploading anything, the experience feels almost invisible. The same goes for a robot that needs to understand spoken commands while still handling navigation and balance. Freeing up bandwidth and compute for those other tasks is a genuine advantage.
Privacy Use Cases That Feel Real
Think about a doctor taking notes during a consultation. Or a lawyer recording a client interview. A journalist capturing sensitive remarks. A therapist session where every word matters and none of it should leave the room. In all those situations the ability to process speech locally removes an entire class of risk.
The model is built for continuous audio. It starts producing text as soon as speech is detected. That design makes it suitable for voice-controlled applications that never send recordings to external servers. For anyone who has watched data-privacy headlines pile up over the past few years, this approach feels refreshing.
Of course no single model solves every privacy concern. But keeping the audio on the device is a solid first step. It shifts the default from “send everything to the cloud” to “process it right here.” That shift alone is worth paying attention to.
How It Fits Into the Broader QVAC Effort
TranscribePsy is not an isolated experiment. It sits inside Tether’s larger QVAC initiative, which aims to run AI applications directly on consumer hardware. Earlier this year the team released an open-source SDK that supports iOS, Android, Windows, macOS and Linux. The toolkit covers text generation, speech, image recognition and translation. It relies on a customized version of llama.cpp called QVAC Fabric and uses peer-to-peer technology for model distribution.
Before the SDK arrived, the group had already integrated BitNet LoRA support. They demonstrated fine-tuning models up to 3.8 billion parameters on devices such as the Pixel 9, Galaxy S25 and iPhone 16. On the iPhone 16 they even tested models as large as 13 billion parameters. GPU acceleration delivered speed gains ranging from two to eleven times over pure CPU runs.
Looking further back, the project released a large synthetic dataset called Genesis I containing 41 billion tokens focused on mathematics, science and technical reasoning. Alongside it came QVAC Workbench, a consumer-facing app that lets people run supported models locally. The pattern is consistent: keep the intelligence close to the user and reduce dependence on centralized infrastructure.
In September the same team showed progress on a peer-to-peer search engine. They distributed the full Wikipedia archive across 100 nodes and demonstrated working search without a central service. The architecture is still evolving toward thousands of interconnected nodes, but the direction is clear.
Four Practical Versions for Different Hardware
The release includes four distinct packages so developers can match the model to their target device.
- A full laptop model supporting 25 European languages
- A reduced language set covering eight common languages
- A mobile GGUF build running through the parakeet.cpp engine inside qvac-fabric-speech
- A Core ML version optimized for Apple’s Neural Engine
Each version is meant to handle continuous audio streams and produce text with minimal delay. The open weights are published under the Apache 2.0 license, making them available for research, education and commercial experimentation.
Inside the wider QVAC Psy family the speech model is expected to work alongside other specialized components such as VisionPsy, TranslatePsy-Nano and MedPsy. Possible future combinations include document analysis, on-device translation, navigation assistance and private note-taking. The October release supplies the speech recognition piece; the rest remains prospective for now.
Benchmark Numbers in Context
Word error rate remains the standard way to judge speech recognition quality. Lower percentages mean fewer mistakes when the system’s output is compared with the original spoken words. TranscribePsy’s results place it competitively among models of similar or larger size.
The research team published the exact evaluation configurations so others can reproduce the numbers. That transparency is welcome. It lets independent researchers verify the claims rather than take them on faith. In a field where benchmark cherry-picking sometimes occurs, releasing the full setup is a quiet but important detail.
When the model was stacked against Qwen3-ASR-0.6B, omniASR-CTC-1B-v2, OWSM-CTC-v4 and Whisper-Medium, it often held its own or pulled ahead on multilingual tests despite the smaller parameter count. Efficiency at this scale is hard to achieve. Seeing a relatively compact model punch above its weight suggests careful training and architecture choices.
What This Means for Everyday Devices
Smart glasses present a particularly interesting case. Audio can arrive over Bluetooth or Wi-Fi, yet the heavy lifting still happens on the glasses or the paired phone. Latency stays low. Battery impact remains manageable because the model is lightweight. For users who want voice notes, real-time captions or discreet translation, the local approach removes friction.
Humanoid robots gain a different benefit. Onboard processors already juggle vision, balance and path planning. Adding cloud-dependent speech recognition creates extra points of failure and extra latency. Running the model locally keeps the conversation loop tight and frees cycles for the physical side of the robot’s work.
Phones, of course, are the most familiar platform. Many of us already use voice input throughout the day. The difference here is that the transcription never leaves the handset. That simple fact changes the risk profile for anyone who handles sensitive material.
The Broader Trend Toward Edge AI
Tether is not the only group working on on-device intelligence, yet its combination of open weights, multi-platform support and explicit privacy messaging feels distinctive. Most large speech systems still assume a cloud connection. Moving the capability onto the device itself aligns with a growing preference for local processing across many AI domains.
I’ve noticed that people tend to accept cloud AI when the benefit feels large and the data feels mundane. The moment the content becomes personal, medical or professional, the calculus shifts. Models like TranscribePsy speak directly to that shift. They offer capability without the usual trade-off of handing over the raw audio.
Of course hardware limits still exist. Not every phone can run the largest language models. Not every pair of glasses has unlimited battery. The 254-million-parameter size appears to be a deliberate compromise that balances accuracy against resource use. In practice that compromise may prove more useful than a higher-accuracy model that drains the battery in minutes.
Open Source and the Path Forward
Because the weights are open under Apache 2.0, developers can inspect, adapt and redistribute the model. That openness lowers the barrier for experimentation. A small team building a specialized medical transcription tool or a privacy-focused journaling app can start from a known baseline rather than training from scratch.
The research group operating under Tether Data continues work on compact models for audio, text, images and video. The long-term vision is a suite of specialized, efficient components that run comfortably on consumer devices. Speech recognition is simply the latest piece to land.
Whether the wider ecosystem adopts these tools remains to be seen. Adoption usually depends on ease of integration, real-world reliability and clear documentation. Early signs look promising: the team has already shipped SDKs, demonstrated multi-device fine-tuning and released evaluation setups. Those steps matter more than any single benchmark score.
Practical Considerations for Developers
Anyone planning to integrate TranscribePsy will want to match the model variant to the target hardware. The full 25-language version makes sense for desktop or laptop applications where storage and memory are less constrained. The eight-language mobile builds trade some coverage for smaller footprints. The Core ML package unlocks the Neural Engine on compatible Apple devices and can deliver the lowest latency on those platforms.
Continuous audio handling is built in, so applications that need live captions or always-on listening can start producing text as soon as speech begins. That design reduces the need for awkward push-to-talk interfaces. At the same time, local processing means the application never has to manage network retries or cloud outages for the transcription step itself.
Memory and disk usage figures released by the team suggest the iPhone-optimized build is especially lean. For developers targeting older or mid-range phones, those efficiency gains could decide whether the feature is viable at all.
Looking Beyond the Launch
The release arrives at a moment when both consumer hardware and user expectations are shifting. Smart glasses are edging closer to mainstream availability. Robots are beginning to appear in controlled commercial settings. And public conversation around data privacy continues to intensify. An open, efficient, offline speech model sits neatly at the intersection of those trends.
I find the combination of performance and principle particularly interesting. The model is fast enough to be useful. It is small enough to run on ordinary devices. And it is framed explicitly as a way for users to keep control of their conversations. That framing may prove as important as the technical numbers.
Whether you are building applications for glasses, robots, phones or something still on the drawing board, the availability of a capable local speech engine expands the design space. You no longer have to assume a permanent network connection or accept the privacy trade-offs that come with cloud transcription. That freedom is rare in today’s AI landscape.
Of course the real test will come in daily use. Benchmarks and controlled tests only tell part of the story. How the model handles accents, background noise, overlapping speakers and domain-specific vocabulary will determine its long-term value. The open release at least gives the community a chance to explore those questions quickly.
In the end, TranscribePsy is less about one company shipping one model and more about a broader idea: useful AI does not have to live in a distant data center. Sometimes the smartest place for it is right there in your pocket, on your face, or inside the machine standing next to you. Keeping the conversation local may turn out to be one of the more practical privacy choices we can make as voice interfaces become part of ordinary life.
