Episode Details

Folk wisdom suggests automated speech recognition fine-tuning stops paying off after fifty hours of audio, but model architecture tells a different story. The fifty-hour mark marks the ceiling for inexpensive language model adjustments, but stronger acoustic adaptation requires hundreds of hours. An exploration of code-mixed datasets, rare-word synthesis, Zipf's law in training data, and simple techniques like weight averaging demonstrates why continuing to fine-tune on personal vocabulary avoids catastrophic forgetting while steadily lowering word error rates. Key sources: • NVIDIA, model card, last modified 2026-08-05, retrieved 2026-11-11. — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 • NVIDIA Perspectives, How much training data do I need to fine-tune a multilingual streaming ASR model?, retrieved 2026-11-11. — https://perspectives.nvidia.com/nemotron-speech/task/faq/how-much-training-data-do-i-need-to-fine-tune-a-multilingual-streaming-asr-model/ • Google, The Gift of Feedback: Improving ASR Model Quality by Learning from User Corrections through Federated Learning, retrieved 2026-11-11. — https://arxiv.org/abs/2310.00141 Episode 6536 — myweirdprompts.com/n/6536 (permanent short link: myweirdprompts.com/295917) Listen on Android: MWP FM Player — myweirdprompts.com/apk