Parakeet TDT v3 adds 25 languages, 3000x speed

Parakeet TDT is the speed king of ASR, effectively rendering Whisper-based pipelines obsolete for high-volume tasks, but it prioritizes "clean" readability over verbatim audio records. Version 2 remains the precision choice for English-only tasks with a 6.05% WER compared to v3's 6.32% WER. v3's primary value lies in its 25-language multilingual engine and improved robustness against non-speech audio hallucinations. Both models aggressively filter "ums" and "uhs," making them unsuitable for legal or clinical verbatim requirements. Achieving 3000x RTFx means one hour of audio is processed in ~1 second, enabling massive-scale transcription at negligible cost.