English-Hebrew code-switching fine-tune of Parakeet released

A Jerusalem-based developer released Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching, a fine-tune of NVIDIA's Parakeet v2 that correctly transcribes Hebrew words in English speech using Latin letters. Code-switching WER dropped from 9.9% to 4.8%, and recall on Hebrew terms unseen in training rose from 29% to 50%. The model ships as .nemo, GGUF, ONNX and sherpa-onnx and runs offline on Android.

• Code-switching WER cut from 9.9% to 4.8% with no English regression
• Unseen Hebrew term recall: 29% for stock v2 vs 50% for the fine-tune
• Training took 23 minutes on one GPU with 120 minutes of audio
• GGUF Q8_0 (730 MB) and Q4_K_M (475 MB) builds available for Android