HuggingFace released TranscribePsy-250M-EU on October 14, 2026, saying it achieves the lowest word error rate of models in the 242M to 370M range on all three benchmarks. It is the company's first multilingual speech recognition model update since the release of Whisper-small in 2023.
HuggingFace reported a word error rate of 8.02% on the Open ASR Leaderboard, measured against models like Whisper-small, which had a 12.34% WER. That compares with 15.31% for omniASR-CTC-300M-v2 and 9.72% for OWSM v4 small.
TranscribePsy-250M-EU is built on a FastConformer-TDT transducer and targets on-device speech recognition for laptops and mobile phones. Availability begins with open weights under Apache 2.0, initially for developers and researchers.
"TranscribePsy was built to remove that trade-off," said Mohammadreza Zolfaghari, lead researcher at Tether AI Research. "The model is small, quantizable, quick to load, and fast enough that transcription finishes well before the user expects it."
The announcement follows the release of Whisper-small and omniASR-CTC-300M-v2, which HuggingFace framed as significant steps in improving on-device speech recognition. HuggingFace did not say how the model will perform on non-European languages, and it raised the open question of how it will scale to more languages.
Source: huggingface
