Inline text normalization for Voice AI agents! Real-time TTS models still struggle with non-standard text. Numbers, dates, currencies, abbreviations - the structured text that shows up in every production voice agent. "$42.50" should sound like "forty-two dollars and fifty cents," not "dollar sign four two period five zero." "03/15/2024" should be "March fifteenth, twenty twenty-four," not "zero three fifteen twenty twenty-four." Async released an industry benchmark to measure how well streaming TTS models handle this. The test: 2,200+ non-standard words across 1,000+ sentences in true streaming conditions with no preprocessing and no text rewriting. Results across four commercial models: • Async Flash v1.0: 81.2% sentence-level accuracy, 88.6% unit-level accuracy • Inworld TTS-1: 49.8% sentence-level, 67.8% unit-level • ElevenLabs Flash v2.5: 40.3% sentence-level, 56.4% unit-level • ElevenLabs Multilingual v2: 37.9% sentence-level, 52.5% unit-level Sentence-level accuracy requires every word in a sentence to be normalized correctly. Unit-level accuracy scores each word independently. Why inline normalization: Traditional TTS pipelines preprocess text before synthesis. Async normalizes during synthesis, improving accuracy for structured text in real-world voice agent contexts. The benchmark includes audio samples where you can hear the normalization differences across models. I've shared the link in the replies!