Key Notes
- Tether released a 254-million-parameter speech recognition family designed to process audio locally without cloud transcription.
- The laptop model supports 25 European languages, while the compact mobile builds cover eight languages and transcribe complete utterances.
- Tether’s reported iPhone benchmark of 1.65 seconds for 10 minutes of audio excludes the initial model-loading and compilation step.
Tether AI Research has released QVAC TranscribePsy, a compact speech-recognition model designed to turn audio into text on local hardware, extending its push into private AI for phones, smart glasses and humanoid robots.
The 254-million-parameter model was announced on October 14. Tether says developers can use it without sending recordings to cloud servers, including processing audio from connected glasses or running recognition on a robot’s onboard computer.
The release provides models and deployment builds, rather than announcing a new pair of glasses or a commercially available robot. It also comes with important differences between its laptop and mobile versions, and between the company’s real-time ambitions and the batch-transcription builds available today.
Local Speech Recognition, With a Smaller Model
The main model card describes a family distilled from Nvidia’s Parakeet-TDT-0.6B-v3. Tether’s researchers reduced the larger model and trained the resulting system using knowledge distillation, a process that transfers capabilities from a larger teacher model into a smaller one.
The broadest release supports 25 European languages, including English, French, German, Polish, Russian, Spanish and Ukrainian. A separate eight-language variant focuses on Dutch, English, French, German, Italian, Polish, Portuguese and Spanish.
Those versions serve different hardware requirements. The 25-language model is published as a PyTorch checkpoint, while the eight-language version also has a quantized GGUF build for phone processors and a Core ML build for Apple’s Neural Engine.
The mobile files are substantially smaller than the full checkpoint. The GGUF model is listed at 305 MB, while the Core ML package totals 212.4 MB. The main 25-language checkpoint occupies about 1.02 GB in its published floating-point format.
That distinction affects deployment: the full language range is not automatically available in the smallest phone build. Developers choosing a version must balance language coverage, memory requirements and the hardware on which recognition will run.
What the iPhone Speed Test Actually Measures
Tether reports that its Core ML build transcribed a single 10-minute recording in 1.65 seconds on an iPhone 17 Pro Max. The benchmark details specify that this result excludes cold start, the initial loading and compilation step.
For that recording, the model card lists a 12.7-second cold start, including the one-time compilation Apple performs when loading the model on a device. The 1.65-second figure therefore measures transcription throughput after preparation, rather than the entire first-use experience.
A different test, covering 160 utterances totaling 40 minutes of audio, took 32.8 seconds with cold start excluded. The researchers explain that short clips require padding to the Neural Engine’s fixed input windows, while a longer recording uses those windows more efficiently.
These are the developers’ reported measurements, not independently reproduced results from CoinScreamer. They demonstrate why the recording format and test conditions matter when comparing a headline speed figure with an everyday voice interaction.
The main model card also reports word error rates of 8.02% on its Open ASR Leaderboard evaluation, 6.58% on Multilingual LibriSpeech and 21.56% on Common Voice. Lower scores indicate fewer transcription errors under the respective test protocols; the three results describe different evaluation sets.
Released Builds Transcribe Complete Utterances
Despite the announcement’s emphasis on near-real-time voice interaction, the released model cards explicitly list batch transcription, rather than streaming, as a limitation. The software transcribes complete utterances, so the release should not be read as a finished, continuously listening assistant.
The mobile documentation also notes that long recordings require careful memory management. Its command-line engine uses overlapping windows for recordings longer than a minute, while application developers must configure their own windowing.
The Core ML version requires Apple hardware with a Neural Engine and iOS 18 or later. The model cards also exclude languages outside their published lists and caution against using the models for safety-critical or high-stakes automated decisions.
The weights and builds are labeled Apache 2.0, with evaluation material published alongside them. The project is an automatic speech-recognition system: converting speech into text is its role, rather than independently understanding a request, choosing an action or controlling a machine.
QVAC’s Wider Push Into Private AI
Tether is positioning local processing as an alternative to uploading personal conversations for transcription. Its announcement identifies smart glasses, robotics and sensitive professional conversations as possible applications, while integrations with its vision, translation and medical-model families remain potential developments.
The release builds on the strategy described in CoinScreamer’s earlier coverage of Tether’s AI applications. That report examined the company’s interest in software that runs on users’ own devices, including environments where internet access is limited.
Our coverage of its robotics investment also explored Tether’s move into autonomous hardware. TranscribePsy adds a speech-processing component to that wider direction; this announcement does not establish that the model has been deployed across a commercial robot fleet.
For developers, the immediate deliverable is a set of downloadable transcription models with documented hardware, language and performance trade-offs. Turning those components into a private voice assistant for glasses or robots will still require the surrounding application, recording controls and device integration.
Disclaimer: CoinScreamer is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and market insights on digital assets and related technologies. NuvexMedia LLC invests in and collaborates with companies across the digital asset, blockchain, and technology sectors. These relationships do not influence CoinScreamer’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2025 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
