nvidia/nemotron-asr-multilingual/asr

{
  "output": "Actually, I'm second guessing myself a single coherent passage might be cleaner and more representative of real world speech performance, which is what benchmarking typically uses. I'll write an engaging flowing piece on something like a journey or a day in a tech forward city. That weaves in all the varied elements naturally, rather than splitting it into sections. I'm deciding to keep the targeted tricky sentences section after all. It'll directly test homophones and number heavy lines that are most prone to errors, so I'll include it as a concise, clearly marked addition at the end now. I'm drafting the full passage, starting with a detailed travel narrative that weaves in dates, times, currency, and specific details by the time we pulled into Waverley Station, my phone had buzzed forty seven times, mostly from doctor Amara Akon Reyes at Neurosynt Technologies fretting over our board presentation. We'd spent three weeks refining a transformer model with one point three billion parameters, and the improvements were stunning. Our word error rate plummeted from fourteen point two percent down to three point eight percent. There's something surreal about watching those number shift like that. They're not just metrics they're the payoff from countless late nights and too much coffee."
}