Async benchmarks streaming TTS normalization

This is vendor marketing, but it lands on a real production pain: voice quality demos hide the places where TTS systems sound careless in actual applications.

• –Async’s methodology is unusually transparent for a vendor benchmark, with downloadable samples, transcriptions, category rules, and aggregate metrics.
• –The strict streaming setup matters because many teams can clean text with preprocessing in batch TTS, but low-latency voice agents often need native handling.
• –The benchmark says Async Flash hit 81.2% sentence accuracy and 88.6% unit accuracy, but LLM-as-judge scoring and vendor-selected categories still deserve skepticism.
• –Developers building voice agents should treat this as a checklist: test prices, dates, URLs, codes, identifiers, and phone numbers before trusting a TTS provider in production.