Tether AI Research announced TranscribePsy on October 14, a 254-million-parameter speech-recognition model in the QVAC ecosystem. The publisher describes support for 25 European languages on a computer and eight on a phone. That difference matters: full-model coverage should not automatically be attributed to a mobile application.
The model uses FastConformer-TDT and was produced by structured pruning of Parakeet-TDT-0.6B-v3 followed by distillation from the original model. The aim is on-device transcription without routinely sending recordings to a server. That can change application privacy and offline behaviour, but depends on the actual integration, not merely downloading weights.
The weights are announced under Apache 2.0. QVAC reports speed and error-rate improvements against its selected comparators under its own protocol and devices; those are not independent measurements across all phones. For dictation, meetings and accessibility, supported languages, real audio, memory requirements and operating conditions matter more than a universal ranking.