Unverified50% confidenceBenchmarkExact time
云端STT服务通常耗时200-500ms，本地部署的Whisper等模型在GPU加速下可压缩至100ms以内
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/14/2026
First Seen
Valid until: 10/12/2026
Sources
状态机语音Agent实战：Pipecat与Vapi的延迟与准确性权衡
redditr/aiagents7/12/2026
Related Claims
Unverified本地部署延迟由三段构成：faster-whisper在GPU上STT约100-300ms、7B参数LLM首Token约200-500ms、TTS合成约100-300ms76% similarUnverified业界领先的STT实时转录延迟约在200-400毫秒，采用Whisper或DeepSpeech等模型72% similarVerified大型模型的Serverless GPU冷启动流程（分配实例、拉取镜像、加载权重）可能耗费30秒乃至数分钟71% similarUnverified云端TTS服务的首字节响应时间通常200-500毫秒，本地部署的轻量TTS模型（如Piper、Kokoro-82M）可将延迟压至50毫秒以内，但音质通常略逊68% similarVerified本地局域网内的指令执行延迟可低至5毫秒以内，远优于云端方案50-300毫秒的往返时延68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/503986API
curl https://kongchang.com/api/v1/knowledge/claims/503986MCP
get_claim(id=503986)