- Trending
- Comments
- Latest
Tether has released an open-source AI model that transcribes speech on the device itself, so recordings never reach a cloud server. The model, called TranscribePsy, converted 10 minutes of audio in 1.65 seconds on an iPhone 17 Pro Max, according to Tether AI Research. It is the latest release from the stablecoin issuer’s QVAC project, announced on October 14.
Unlike speech-recognition systems that depend on remote servers, TranscribePsy is designed to process audio locally. This approach reduces the need to upload conversations to external platforms, potentially limiting exposure of sensitive recordings during transcription.
The distinction could be significant as voice-enabled technology becomes more common across mobile devices, wearable electronics and robotics. Smart glasses, for example, can receive audio through Bluetooth or Wi-Fi and process it with minimal delay. Humanoid robots can run the software on onboard processors, leaving other computing resources available for tasks such as navigation, visual recognition and movement.
Tether chief executive Paolo Ardoino said the company wants users to retain greater control over their personal information as AI-powered devices become more widespread.
“The future of AI shouldn’t mean giving centralized platforms a front-row seat to our personal lives,” Ardoino said.
The company presents TranscribePsy as one way to support that objective by keeping speech processing on personal devices rather than relying on centralized services.
However, local processing alone does not guarantee complete privacy. The security of a finished application will also depend on how developers handle stored recordings, generated transcripts, permissions and any information shared with other services.
Potential applications include transcribing interviews, documenting consultations and generating private notes. These uses remain dependent on how developers integrate the model into their products.
Tether’s published evaluations indicate that TranscribePsy performed competitively against several larger speech-recognition systems.
The company reported a word error rate of 8.02% on the Open ASR Leaderboard and 6.58% on Multilingual LibriSpeech. On Common Voice, the model recorded a word error rate of 21.56%.
Word error rate measures the difference between a system’s transcription and the original spoken words. Lower scores indicate fewer transcription errors, although results can vary depending on language, recording conditions, accents and evaluation methodology.
According to Tether’s research team, the model outperformed several competing systems within its parameter category. The comparisons included Qwen3-ASR-0.6B, omniASR-CTC-1B-v2, OWSM-CTC-v4 and Whisper-Medium.
The company also reported that the model outperformed several systems three to four times its size on the Multilingual LibriSpeech benchmark.
Smartphone testing provided another measure of performance. Across six devices, Tether said the model ran 2.6 times faster than the next-fastest system and 5.7 times faster than Whisper-small when operating on phone CPUs.
An iPhone-optimized version also required less than half the disk space and 1.7 times less memory than a competing Core ML model included in the company’s evaluation.
These figures are based on Tether’s reported tests and should be understood in that context. The publication of evaluation configurations gives other researchers an opportunity to reproduce the results and assess performance under comparable conditions.
TranscribePsy is available in four versions intended for different hardware configurations and language requirements.
The laptop version supports 25 European languages, while a smaller language configuration focuses on eight commonly used languages. The mobile release includes a GGUF version that operates through the parakeet.cpp engine in qvac-fabric-speech.
For compatible Apple devices, a separate Core ML build is optimized for Apple’s Neural Engine. This allows supported iPhones to use dedicated hardware for speech-recognition tasks instead of relying exclusively on general-purpose processing.
The software is designed to handle continuous audio and generate text shortly after speech is detected. That capability could help developers build voice-controlled applications that respond quickly without transmitting recordings to external servers.
Tether has released the model under the Apache 2.0 license, making its weights, mobile builds, Apple Silicon Core ML versions and evaluation configurations available for development and research.
Within the QVAC Psy ecosystem, the speech-recognition model is intended to complement other specialized tools, including VisionPsy, TranslatePsy-Nano and MedPsy. The broader ecosystem could eventually support applications combining audio transcription, image analysis, translation and other AI functions.
Tether has identified document analysis, navigation and private note-taking as potential applications for future integrations. These are prospective uses rather than features established by the speech-recognition release alone.
The launch follows several releases under Tether’s QVAC initiative in 2026, reflecting the company’s continued investment in AI systems that can run on consumer hardware.
In April, Tether introduced the QVAC software development kit, an open-source toolkit designed to support locally run AI applications across iOS, Android, Windows, macOS and Linux. Its capabilities cover areas including text generation, speech processing, image recognition and translation.
The toolkit uses QVAC Fabric, a customized version of llama.cpp, alongside Holepunch technology for peer-to-peer model distribution and delegated processing.
In March, Tether integrated BitNet LoRA into QVAC Fabric to support the training and operation of AI models on consumer devices. The company said it had fine-tuned models containing up to 3.8 billion parameters on devices including the Pixel 9, Galaxy S25 and iPhone 16.
Tether also reported tests involving models with up to 13 billion parameters on the iPhone 16, while GPU processing delivered speed improvements ranging from two to 11 times compared with CPU baselines in its evaluations.
Beyond speech and language models, the company has explored decentralized computing infrastructure. In September, Ardoino disclosed progress on a peer-to-peer search engine designed to operate across distributed nodes. The team had stored the complete Wikipedia archive across 100 nodes and demonstrated search functionality without depending on a centralized search service.
The search engine remains under development, with work continuing on an architecture intended to support thousands of interconnected nodes.
Earlier projects included the QVAC Genesis I dataset, released in October 2025 with 41 billion synthetic tokens intended for training models in mathematics, science and technical reasoning. Tether also introduced QVAC Workbench, a consumer application that allows users to run supported AI models locally.
With TranscribePsy, Tether is extending this approach to speech recognition, a capability that could become increasingly relevant as voice interfaces spread across phones, wearable devices and robotics.
The release gives developers access to a model designed around local processing, multilingual transcription and comparatively modest hardware requirements. Its longer-term impact will depend on independent performance testing, practical integration and whether developers can translate the company’s privacy-focused approach into reliable applications.
Copyright © 2025 - The Bit Gazette.
