Unverified50% confidenceFactExact time
非自回归模型（如FastSpeech系列、VITS）并行生成所有帧，通常比自回归模型快10-100倍但表现力受限
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
H3语音模型全精度权重发布：表现力惊艳但长句声音漂移待解
Related Claims
Unverified非自回归语音模型通常比自回归模型快10-100倍，但表现力受限于预测器准确性67% similarUnverified顶级模型每次调用的算力消耗可能是轻量模型的数十倍63% similarUnverifiedDuring the Decode phase of large models, because each step depends on the result of the previous step and can only be executed serially, hardware utilization is typically below 10%.62% similarUnverifiedDifferent hyperparameters such as system prompt, temperature, maximum token count, and number of retries can cause the evaluation results of the same model to differ by more than 10 percentage points.62% similarUnverified效率型模型以远低于旗舰模型的推理成本（通常低5-10倍）提供接近旗舰模型80-90%的能力表现61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/687641API
curl https://kongchang.com/api/v1/knowledge/claims/687641MCP
get_claim(id=687641)