Latest / Logosofia no Dia a Dia
Why Your TTS Model Nails "Shabbat" but Not "Keren Hishtalmut
Ever notice how your text-to-speech model nails "Shabbat" but completely butchers "Keren Hishtalmut"? It's not a bug — it's a consequence of how training data is distributed. This episode unpacks the tokenization mechanics, the loanword threshold, and the Zipfian distribution problem that leaves niche multilingual vocabulary stranded. We explore why fine-tuning isn't practical for proprietary APIs, and walk through the current workarounds — from SSML phoneme overrides to phonetic spelling hacks — that can rescue your code-switched content today. Episode 2982 — myweirdprompts.com/n/2982…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.