any solutions for sending dtmf tones faster? in situations where the IVR menu has a very short window for sending a dtmf

Thanks for reaching out, Ed. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

i just mean the amount of time it takes the pipeline STT to process the IVR menu speech, and the llm to call a tool that uses publish_dtmf takes too long

Thanks — I’m responding about reducing end-to-end latency so your agent can send DTMF inside very short IVR windows. Quick summary of what’s causing the delay: STT endpointing and model choice add transcription latency, LLM time-to-first-token (and any tool-call orchestration) adds more, and the DTMF publish step is performed only after the LLM triggers it. Practical ways to reduce that overall time • Use streaming STT and tune endpointing: enable streaming partial transcripts and reduce end-of-utterance silence thresholds so the agent gets text earlier. • Start LLM work from partial transcripts: emit partial transcripts to the LLM (preemptive generation) so the model can begin deciding before the full utterance finishes. • Choose faster models or lower-latency endpoints: pick STT/LLM models with lower TTFT or use nearer inference regions to cut network latency. • Parallelize stages: overlap STT → LLM → publish_dtmf instead of strictly waiting for each stage to fully finish. • Short-circuit decision logic: add lightweight, deterministic rules (e.g., regex or keyword matcher) that can call

publish_dtmf

immediately for simple IVR prompts without a full LLM pass. • Instrument & measure: use agent observability to measure STT, LLM, and tool call durations and find the biggest bottleneck to target. Trade-offs to be aware of • Preemptive/partial transcription and faster models can reduce accuracy; you may need retries or fallbacks. • Custom short-circuit logic reduces LLM latency but increases maintenance for IVR edge cases. One quick clarifying question: are you able to stream partial STT results into the LLM today, or is your pipeline currently waiting for a final transcript before calling the model? Sources: Understand and Improve Agent Latency | LiveKit | Voice Agent Architecture: STT, LLM, and TTS Pipelines Explained | LiveKit | Understand and Improve Agent Latency | LiveKit | send_dtmf_events | LiveKit Documentation | Sequential Pipeline Architecture for Voice Agents | LiveKit, with interactive elements