Fine-Tuning Parakeet for Hebrew and Your Own Jargon
Daniel runs NVIDIA's Parakeet V2 on Android for dictation and wants two things: a Hebrew fine-tune, and a personal-vocabulary tune to catch tech jargon and Hebrew loanwords like "makolet." We dig into what Parakeet actually is under the hood — FastConformer encoders with CTC, RNN-T, and TDT decoders, shipped through NVIDIA NeMo — and why its transducer architecture sidesteps the hallucination loops Whisper is famous for. Then we separate the two questions: personal vocabulary turns out to have a lightweight, already-shipped answer in contextual biasing and hotwords, while Hebrew is a genuine…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.