# Tether Unveils On-Device Speech Recognition Model for Phones and Robotics QVAC TranscribePsy enables fast, private multilingual voice transcription without cloud processing. - **Source:** [INCRYPTED](https://brifing.si/source/7cd4625d-3f08-4cc1-983d-4c7d991839ba) - **Published:** 2026-10-14 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/f73f01d3-51b0-462b-be9f-1a6e1846c223 ## Local Multilingual Speech Recognition Tether AI Research has introduced QVAC TranscribePsy, a compact speech recognition model running entirely on-device without cloud connectivity. Containing 254 million parameters, the architecture targets smartphones, smart glasses, and humanoid robots. Its primary goal is ensuring total conversation privacy while delivering near real-time edge processing. ## Hardware Optimization and Speed The model features deep hardware optimizations across diverse hardware ecosystems. When running via Apple Core ML on an iPhone 17 Pro Max, it transcribes 10 minutes of audio in just 1.65 seconds. Tether also released dedicated GGUF format builds to support mobile and embedded chipsets. ## Competitive Word Error Benchmarks In standard evaluation metrics, TranscribePsy scored an 8.02% Word Error Rate on the Open ASR Leaderboard and 21.56% on Common Voice. On Multilingual LibriSpeech, it recorded 6.58% WER, outperforming several competing systems that maintain three to four times higher parameter counts across equivalent benchmarks. ## Smartphone Inference Performance Gains Testing across six distinct smartphones revealed that TranscribePsy operates 2.6 times faster than its nearest edge competitor. Furthermore, it runs 5.7 times faster than Whisper-small on mobile processors. The laptop release supports 25 European languages, while the optimized mobile build handles eight distinct languages. ## Confidentiality for Regulated Sectors By eliminating server-side audio transmissions, the architecture addresses strict data compliance barriers. Tether highlights practical utility in medical consultations, confidential legal depositions, therapy sessions, and investigative journalism. CEO Paolo Ardoino emphasized that advanced AI systems must not grant centralized cloud platforms front-row access to sensitive private user discussions. ## Broader QVAC Ecosystem Synergy TranscribePsy integrates directly into Tether's expanding on-device QVAC suite. Future updates plan deep interoperability with models like VisionPsy, TranslatePsy-Nano, and MedPsy for multi-modal document analysis and navigation. The model follows the June 2026 release of QVAC SDK 0.13.0, which added robotics and brain-computer interface support. ## Synthetic Training Data Foundations The model builds upon Tether's prior AI data initiatives launched in December 2025 with QVAC Genesis II. That synthetic dataset grew by 107 billion tokens to reach 148 billion tokens across 19 educational domains, establishing foundational training pipelines for specialized edge models across the ecosystem. ## Open Weights and Conglomerate Strategy Published under the Apache 2.0 license, TranscribePsy provides open model weights and testing configurations for researchers. This release aligns with Tether's broader transformation beyond digital assets into a diversified technology conglomerate, as affirmed by Paolo Ardoino following recent institutional reserve audit engagements with auditing firm KPMG. ## Sources covering this story - https://incrypted.com/tether-predstavyla-yy-model-dlya-raspoznavanyya-rechy-na-smartfonakh-umnykh-ochkakh-y-robotakh/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._