Unverified50% confidenceFactExact time
现代 ASR 模型在解码每个词时会计算概率分布，识别越有把握置信度分数越接近 1.0，可通过阈值将低置信度词标记为不确定
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe credible interval in Bayesian methods can be directly interpreted as the probability that a parameter falls within the range, providing a more intuitive quantification of uncertainty than the frequentist confidence interval.72% similarUnverifiedASR转录文字稿的准确率上限直接决定下游LLM语义分析的质量上限，转录错误会降低大模型对重复片段、气口位置的判断精度70% similarUnverified每增加一条有效约束等价于在模型的输出概率分布上施加一个条件概率过滤器，将生成空间收窄为符合约束的子集68% similarUnverified投机解码通过小模型预测候选Token、大模型批量验证提升吞吐量；量化技术将权重从FP16压缩至INT8/INT4降低显存占用；MoE架构通过稀疏激活降低单次推理计算量67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/794364API
curl https://kongchang.com/api/v1/knowledge/claims/794364MCP
get_claim(id=794364)