Episode Details

Scoring every word in a vocabulary seems computationally ruinous, but language models and speech recognizers compute every token's probability simultaneously in a single matrix multiplication. Softmax merely normalizes raw scores, while the real cost scales with vocabulary size rather than repeated model evaluations. Examining architectures like NVIDIA's Parakeet reveals how token-and-duration transducers speed up transcription by predicting durations to skip redundant audio frames, paired with greedy argmax decoding. Key sources: • NVIDIA, Hugging Face model card, created 2025-04-15, last modified 2026-06-29, retrieved 2026-11-12. — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2 • NVIDIA, Hugging Face model card, created 2025-08-04, last modified 2026-08-05, retrieved 2026-11-12. — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 • Xu et al., Efficient Sequence Transduction by Jointly Predicting Tokens and Durations (TDT), v2 revised 2023-05-29, retrieved 2026-11-12. — https://arxiv.org/abs/2304.06795 Episode 6538 — myweirdprompts.com/n/6538 (permanent short link: myweirdprompts.com/766771) Listen on Android: MWP FM Player — myweirdprompts.com/apk