Before waveform generation, the input text is processed via a "wiseguy lexicon" that applies phonological rules:
: Studies on accent-based TTS highlight how specific regional dialects (like the New York/New Jersey "mobster" inflection) are synthesized using Recurrent Neural Networks to transfer speech patterns between accents.
Modern systems like VITS (Variational Inference Text-to-Speech) allow for "style transfer." A developer can input text and apply a "style vector" derived from a sample of an angry or whispering speaker. For a Wiseguy voice, the system must handle Code-Switching . A convincing mobster character often switches between a polite, high-pitched "business" tone and a low, gravelly "threat" tone within a single paragraph. Traditional TTS struggles to switch emotional states mid-sentence without introducing artifacts; modern end-to-end models are beginning to solve this by conditioning the model on "speaker embeddings" that define emotional state.