One thing people building voice AI often overlook is how much information gets lost before the LLM even receives a transcript. Think about someone learning English. They'll pause mid sentence, pronounce words incorrectly, repeat themselves, or make grammatical mistakes. Now imagine your STT model quietly cleans all that up. The transcript might look great, but an AI tutor relying on it could completely miss what the student needs help with. That's why I found this benchmark from Fluently interesting. They tested real time speech recognition on low fluency English spoken by Chinese, Japanese, Korean, Thai and Vietnamese learners. And the differences between models are pretty significant. If you're building anything with voice AI, here's what I'd test before choosing a transcription model: - How well does it handle your actual user's accents? - Does it preserve hesitations, repetitions and grammatical mistakes, or quietly clean them up? - How often does it get important words wrong, especially negatives, names and numbers? - How well does it handle pauses, interruptions and spontaneous conversations? A low word error rate is useful, but it doesn't tell you everything about how well the model will work in your application. So we need to test speech recognition on the people who'll actually use your product, similar to the example of Dessa that I shared few days back.