Tether has introduced a 254 million parameter speech recognition model capable of transcribing spoken dialogue directly on smartphones, smart glasses, and humanoid robots. Audio is processed locally without being transmitted to cloud servers.
Tether AI Research revealed in an October 14 update that its open-source tool, QVAC TranscribePsy, accommodates multilingual voice translation while being tailored for low-power hardware systems.
The firm stated that its software transcribes a ten-minute voice recording in just 1.65 seconds on an iPhone 17 Pro Max, leveraging system speed enabled by Apple’s Core ML framework.
TranscribePsy belongs to Tether’s QVAC AI initiative, an effort targeting local artificial intelligence execution. Developer access to offline speech recognition capabilities is enabled by this rollout.
In smart glasses, the system handles incoming Bluetooth or Wi-Fi audio with negligible latency. Humanoid robots execute the software via onboard processors, leaving core capacity reserved for critical functions like navigation, visual tracking, and physical motion.
Tether noted that edge processing offers clear benefits during medical consultations, legal depositions, therapy sessions, and investigative journalism, where sensitive details are often recorded.
Tether’s TranscribePsy Model Delivers Strong Benchmark Performance
According to company benchmark results, TranscribePsy achieved an 8.02% word error rate on the Open ASR Leaderboard, with a 6.58% score recorded on Multilingual LibriSpeech.
The model logged a 21.56% word error rate on Common Voice, with superior performance claimed by Tether over competing systems within the same parameter range.
Evaluators stacked the 254 million parameter model up against heavier competitors, including Qwen3-ASR-0.6B, omniASR-CTC-1B-v2, OWSM-CTC-v4, and whisper-medium, which was also tested.
Despite its smaller parameter count, TranscribePsy outpaced multiple systems three to four times its size on the Multilingual LibriSpeech benchmark, according to findings reported by the research team.
Across six smartphone benchmarks, the model ran 2.6 times faster than its nearest rival and 5.7 times faster than Whisper-small, with top speeds achieved when executed on phone CPUs.
The iPhone-optimized build needed under half the storage space and 1.7 times less RAM than a competing Core ML variant, with performance benchmarked alongside rival models in the evaluation.
Accuracy in speech recognition was measured through word error rates, comparing a system’s output directly against original spoken audio, where lower scores indicate that fewer mistakes were made.
Tether published its evaluation setups along with the model code, allowing identical results to be reproduced by outside researchers.
Speech Recognition Software Adds Support for 25 European Languages
Tether introduced TranscribePsy across four distinct builds, each tailored to handle unique hardware configurations and language demands while four specialized editions were made available.
One laptop edition supports 25 European tongues, while a streamlined language setup targets eight widely spoken languages, with tailored builds offered across both options.
For mobile hardware, Tether supplies a GGUF format that functions via the parakeet.cpp engine, where qvac-fabric-speech is utilized for execution.
A standalone Core ML edition targets Apple’s Neural Engine, enabling compatible iPhones to handle speech transcription when dedicated hardware is utilized.
The models manage continuous audio streams, generating text moments after speech occurs. Tether noted that the framework can power voice-activated tools, with user recordings processed without being sent to external cloud infrastructure.
According to company representatives, local speech processing minimizes the necessity of transmitting sensitive conversations, where personal data is often contained.
Tether CEO Paolo Ardoino emphasized that the firm aims to keep control over private conversations in the hands of users as voice-activated technology spreads, with data privacy prioritized across devices.
“The future of AI shouldn’t mean giving centralized platforms a front-row seat to our personal lives,”
Ardoino said.
He highlighted TranscribePsy as a prime model of speech recognition tech running seamlessly on personal hardware, where full independence from centralized infrastructure is maintained.
Tether Broadens Its QVAC AI Development Toolkit
TranscribePsy arrives on the heels of multiple software rollouts under Tether’s QVAC initiative throughout 2026, with new privacy-focused tools continuously introduced.
In April, Tether debuted the QVAC SDK, an open-source development toolkit designed to power local AI applications across smartphones and desktops, with accessible software assets provided to developers.
As previously reported by , the toolkit supports iOS, Android, Windows, macOS, and Linux, with features spanning text generation, speech processing, image recognition, and translation.
The software leverages QVAC Fabric—a tailored build of llama.cpp—alongside Holepunch technology to handle peer-to-peer model distribution, where task assignments are delegated across the network.
Prior to the SDK launch, Tether integrated BitNet LoRA into QVAC Fabric back in March, aiding the training and execution of massive AI models on consumer-grade hardware where resource efficiency was prioritized.
At that point, the company reported fine-tuning models holding up to 3.8 billion parameters on smartphones like the Pixel 9, Galaxy S25, and iPhone 16, where optimal local efficiency was achieved.
Evaluation trials assessed models scaling up to 13 billion parameters on the iPhone 16, while GPU acceleration yielded speed gains spanning two to 11 times over CPU baselines, with benchmark metrics recorded across internal tests.
Tether’s ongoing exploration into decentralized computing extends well past specialized AI architectures, where broader architectural goals are actively pursued.
In September, Ardoino revealed significant headway on a peer-to-peer search engine built to run across distributed nodes, where decentralized network protocols are utilized.
The team successfully archived the entire Wikipedia database across a 100-node setup, demonstrating full search capabilities where central servers were completely bypassed.
Development on the search engine continues as engineers refine an architecture designed to link thousands of interconnected nodes, where widespread network stability is actively pursued.
Tether’s preceding AI initiatives encompassed the QVAC Genesis I dataset, rolled out in October 2025 featuring 41 billion synthetic tokens to train models in mathematics, science, and technical reasoning, where specialized benchmark datasets were established.
Alongside the dataset, the firm launched QVAC Workbench, a consumer-facing application enabling users to execute supported AI models locally, where full data autonomy is retained.
TranscribePsy Launches With Open Model Weights
The latest TranscribePsy release is available under the Apache 2.0 open-source license, with Tether highlighting research and educational applications, where open academic collaboration is encouraged.
Software creators can access the model weights, mobile-optimized GGUF builds, Apple Silicon Core ML variants, and evaluation frameworks, where open-source developer access is fully enabled.
Within the QVAC Psy framework, the speech recognition engine pairs seamlessly with other domain-specific models, including VisionPsy, TranslatePsy-Nano, and MedPsy, where multi-modal AI workflows are actively facilitated.
Tether highlighted document analysis, translation, navigation, and private note-taking as prospective use cases for future integrations, where expanded cross-functional features are envisioned. While the October release delivers the core speech recognition engine, these combined applications remain forward-looking concepts.
The research team at Tether Data, S.A. de C.V. noted that its ongoing initiative encompasses lightweight models designed to handle audio, text, images, and video on everyday consumer hardware, where localized processing performance is continuously optimized.
