Episode Details

Daniel wants to bend Parakeet to his will: strip the ums, catch niche words, and handle the Hebrew he sprinkles into his English. He's done this before with Whisper, so he knows the shape of the job — but Parakeet is a different animal. This episode walks through what Parakeet TDT 0.6B v3 actually is, why a small model is a better inference candidate but not an easier fine-tune, and the real workflow: NVIDIA NeMo on a rented GPU, a JSON manifest, two learning-rate knobs, then a GGUF conversion and a sideload back to the phone. We dig into the disfluency problem (there's no published Parakeet recipe, but a Whisper LoRA shows the trick: the lesson lives in the target transcript, not in a rule), the Hebrew code-switching problem (why the obvious Hebrew Parakeet model is the wrong tool, and what a Hebrish Whisper fine-tune got right with just 500 phrases), and the tokenizer-extension mechanics that matter if you ever need to teach the model new words without forgetting the old ones. Key sources: • NeMo Fine-Tuning docs — https://docs.nvidia.com/nemo/speech/nightly/asr/fine_tuning.html • NeMo Word Boosting (GPU-PB, CTC-WS) — https://docs.nvidia.com/nemo/speech/nightly/asr/asr_customization/word_boosting.html • parakeet.cpp repo (v0.5.0, 2026-08-01) — https://github.com/mudler/parakeet.cpp Episode 5868 — myweirdprompts.com/n/5868 (permanent short link: myweirdprompts.com/766682)