When Your TTS Model Eats the Numbers

Text normalization is the invisible layer that turns written text into something a TTS model can actually speak: "1,234" becomes "one thousand two hundred thirty-four," "Dr." becomes "Doctor," and "2026" stops being a string of digits. It's also the layer this podcast doesn't have, which is why acronyms and numbers have been hand-patched for months. This episode walks through the vocabulary — text normalization for TTS, inverse text normalization and punctuation restoration for ASR — then digs into the documented failure modes of Chatterbox, the open-source TTS engine behind the show, whose entire preprocessing pipeline is a single punctuation cleanup function. Numbers produce unknown tokens, internal capitals trigger emphasis behavior that cuts audio short, and acronyms like S&ED fall through entirely. The conversation covers NVIDIA NeMo, WeTextProcessing, and Apple's prompt-based PolyNorm, whose benchmark results show a prompted GPT-4o beating a rule-based production baseline across eight languages. The throughline: normalizing text for a specific TTS model requires knowing which model you're targeting. Episode #051746 — open it directly at myweirdprompts.com/051746