AirCaps AI launched A5S Async v1, an API that transcribes recorded meetings, calls, interviews, and media and marks who said each word, including stretches where people talk at the same time. It is meant for teams building notetakers, captions, and call review, where the transcript has to carry both the words and the speaker. On launch day VoiceArena ranked it first of 15 systems on Diarization-Bench, a set of 139 natural English conversations with two to eight speakers. The next-best system on that board, NVIDIA's Nemotron 3.5 Transcribe, makes 36% more errors. On five public meeting benchmarks, in English and Mandarin, the model has the lowest error on each set. Across those five, the next-best model makes 41% more errors. AirCaps published the harness used for that comparison, including dataset recipes, provider settings, and raw API responses. An hour of audio comes back in a median 9.8 seconds at $0.18 per audio hour, with speaker labels, word timestamps, and SRT or VTT subtitles included. It supports English, Spanish, French, German, Portuguese, Hindi, Japanese, Korean, and Chinese. New accounts include five free hours. Madhav Lavakare and Nirbhay Narang founded AirCaps, a Y Combinator Fall 2025 company in San Francisco. The team started with live captions for in-person conversations and now offers speech models other products can call.