AINetEase Youdao's speech recognition reportedly shares one foundation for offline and streaming
A user describes an open-source design pairing an audio encoder with a large language model, using configurable chunk sizes to balance response time and recognition quality.
TLDR
A user says NetEase Youdao's open-source speech-recognition system uses one audio-encoder-plus-LLM foundation for both offline and streaming recognition. They argue that configurable chunk sizes allow tuning for latency and quality without training and maintaining two separate models, and claim there is no accuracy penalty for offline use. Their follow-up highlights domain-aware recognition and treating context as a core input.
Combined views
657
1 Source, first seen 1d ago
NetEase Youdao's speech recognition reportedly shares one foundation for offline and streaming
A user describes an open-source design pairing an audio encoder with a large language model, using configurable chunk sizes to balance response time and recognition quality.
TLDR
A user says NetEase Youdao's open-source speech-recognition system uses one audio-encoder-plus-LLM foundation for both offline and streaming recognition. They argue that configurable chunk sizes allow tuning for latency and quality without training and maintaining two separate models, and claim there is no accuracy penalty for offline use. Their follow-up highlights domain-aware recognition and treating context as a core input.