What TranscribePsy is
Tether AI Research released TranscribePsy on October 14, 2026. It is a speech recognition model with 254 million parameters. According to Tether:
- There are two versions: one with 25 European languages and one with 8. Portuguese is in both.
- On the phones tested, the 8-language version used 530 to 542 MiB of memory.
- It was trained on about 564,000 hours of speech, from NVIDIA's Parakeet-TDT-0.6B-v3.
- On Multilingual LibriSpeech, it makes fewer errors than models of a similar size, such as Whisper-small.
Who it is for
- Consultations and interviews where the audio must not leave the device.
- Captions for recorded talks and videos, without a network.
- Voice notes and meeting minutes on the phone itself.
Why through Rota Nacional
- We bring the model to your organization's app or devices, or we run it on Rota's workers, on the same API.
- We measure the quality in Brazilian Portuguese with your recordings before use.
How it works
You describe the case
The type of recording, the devices and the volume.
We test with your samples
Tether has not published per-language figures yet. So we measure Brazilian Portuguese with your recordings.
We connect it to your access
The model joins your app or the API you already use.
Good to know
- Weights under the Apache 2.0 license. Tether says it released them for research and education; we confirm commercial use with Tether before deployment.
- The source does not say whether the Portuguese is European or Brazilian.
- It does not transcribe live yet: a recording is transcribed when it ends.
- On-demand deployment: the model goes live when an organization asks for it.