Rewriting TTS Inference Economics: Lightning V2 on Tenstorrent Achieves 4x Lower Cost Than NVIDIA L40S

Ranjith M. S., Akshat Mandloi, Sudarshan Kamath

Text-to-Speech (TTS) models are significantly more numerically fragile than Large Language Models (LLMs) due to their continuous waveform generation and perceptual sensitivity to s…