NetEase Youdao open-sources R2T2 streaming ASR

LLM-backed streaming ASR with runtime context injection is the missing link for enterprise voice agents that constantly stumble over specialized jargon and unstable transcriptions.

• –**Append-only streaming reliability:** Waiting for confident acoustic evidence before committing tokens ensures downstream agents do not trigger actions based on speculative partial text that later gets rewritten.
• –**Dynamic runtime hotword injection:** Utilizing an LLM decoder allows developers to feed context like meeting participants, product databases, or codebase entities into prompt context without retraining or fine-tuning weights.
• –**Production-ready serving options:** Native support for high-performance inference backends like vLLM makes deploying low-latency, self-hosted streaming ASR scalable and cost-effective compared to proprietary voice APIs.
• –**Unified dual-mode architecture:** Consolidating offline batch transcription and real-time streaming into one architecture reduces operational overhead and engineering complexity for multimodal voice stacks.