New Voice AI Handles Greek and English in Noisy Rooms

Researchers spent a multi-month engineering program building Sophea, a production bilingual Greek-English automatic speech recognition system, and evaluated it against nine production gates covering Greek and English word error rate, language identification, and hallucinations on non-speech audio. Across twenty-three training iterations and two model architectures, no training-data composition passed all nine gates at once. Meeting the Greek noisy-environment target required about 1,500 steps of dense domain exposure, while preserving English language identification tolerated only about 250 steps, or about 1,250 with a rebalanced mix that reduced Greek accuracy. One fix that worked: a three-model ROVER ensemble raised gate coverage from 4-7 of 9 for individual models to 9 of 9 and cut overlapping-speech word error rate from 53.35 percent to 37.87 percent, a 29 percent relative improvement. Separately, a learned per-clip arbiter over two models is listed as sophea/asr-k1 (preview) on the public Open ASR Leaderboard, with 4.26 percent average word error rate across eight public English test sets, and reaches 25.88 percent word error rate on live Greek noisy-environment traffic. No model weights or training data are released; the paper reports methodology and quantitative results only.