Streaming ASR
Streaming automatic speech recognition, which employs chunk-wise decoding to output partial hypotheses in real time. Constrained by latency, it cannot use full future context and is prone to error propagation caused by the fixation of early errors.
Timeline (last 90 days)
Mistral's streaming model treats target latency as a model parameter that users can flexibly control based on their scenario, and can even run multiple streams with different latencies simultaneously (a fast stream for immediate display and a slow stream for subsequent correction).
Streaming models have more failure modes than batch models and often hallucinate additional speakers out of thin air, which is a direct consequence of insufficient context.
All Facts (2)
Mistral's streaming model treats target latency as a model parameter that users can flexibly control based on their scenario, and can even run multiple streams with different latencies simultaneously (a fast stream for immediate display and a slow stream for subsequent correction).
50%UnverifiedStreaming models have more failure modes than batch models and often hallucinate additional speakers out of thin air, which is a direct consequence of insufficient context.
50%