Factorized Delayed Streams Modeling for LLM-based Streaming ASR
Factorized DSM is proposed, which separates the waiting probability for ASR-specific tokens from the text prediction space and allows the large-vocabulary softmax to be skipped on waiting steps and greatly reduces GPU memory use while maintaining similar training throughput.
Tatsunari Takagi, Kai Washizaki, Atsushi Kojima et al.
· 0 citations