Ten minutes of recorded speech turned into text in seconds sounds useful. Before we applaud, though, we would like to see the transcript.
Tether, the company behind the USDT stablecoin, has released a speech-to-text model called QVAC TranscribePsy. Developers can use it to build applications that process recordings on the device, without sending the audio to a remote server.
Strip away the smart glasses, humanoid robots, and sweeping promises about privacy, and that is the announcement: software that transcribes recordings on your own device. Keeping conversations off someone else’s server is a real benefit. It tells us nothing about how accurately the software will transcribe them.
Speed is only half the story
What matters is how much editing the transcript needs. A real meeting rarely follows a neat script. People mumble, interrupt each other, and mention unfamiliar names. Getting one number wrong can matter more than missing ten commas.
Tether does publish accuracy results. In its tests, the 25-language model records a word error rate of 8.02% on Open ASR and 21.56% on Common Voice. The results come from different datasets, so they should not be read as interchangeable measures of accuracy.
Those figures deserve attention. They do not tell us how much editing your next meeting transcript will need. We have not independently tested the model. If a transcript needs extensive corrections, faster processing simply gets you to that work sooner.
About those 1.65 seconds
Tether reports that the model transcribes ten minutes of audio in 1.65 seconds on an iPhone 17 Pro Max. Its technical documentation includes an important qualification: the figure excludes a 12.7-second cold start, during which the model is prepared to run.
That startup delay does not apply to every recording. The 1.65-second figure describes processing once the model is ready, not the entire wait from a cold launch. It is also a result from a specific phone, not a promise about every handset.
Live transcription is still missing
Despite all the talk of real-time interaction, the official technical FAQ says this version processes a recording or utterance once it ends. Streaming recognition is still being developed.
The ready-made phone builds support eight languages. The 25-language release, which includes Russian, is a separate version for computers and servers. That broader language support does not carry over to the available mobile builds.
Now try a real conversation
There is something worth trying here: downloadable models, local processing, published benchmarks, and an Apache 2.0 license listed in the model card. Developers have actual software to work with.
Now try a conversation with background noise and people interrupting each other. Publish the transcript, count the corrections, and compare the result with an existing tool using the same recording. Speech-to-text software earns its place by saving work. The robots can wait.