Tencent Hunyuan Releases Hy ASR 3.0 with ~3% Word Error Rate

Tencent Hunyuan launched Hy ASR 3.0 preview, a new speech recognition model. It uses the latest LLM Hy3 to understand context, not just transcribe words. The model improves accuracy in general recognition, dialects, and noisy environments. Word error rates for Mandarin, English, and Cantonese are 3.34%, 2.62%, and 3.12%, respectively, leading open benchmarks.

Hy ASR 3.0 is now available via Tencent Cloud API for uses like customer service and content understanding. Tencent's Yuanbao app has integrated it, letting users speak for dialect recognition and context correction, free. WorkBuddy will soon add it too.