Voice agents represent the frontier of AI interaction - humans speaking naturally with AI systems.

Voice Agents Voice agents represent the frontier of AI interaction humans speaking naturally with AI systems. The challenge isn't just speech recognition and synthesis, it's achieving natural conversation flow with sub 800ms latency while handling interruptions, background noise, and emotional nuance. This skill covers two architectures: speech to speech (OpenAI Realtime API, one low latency architecture option) and pipeline (STT→LLM→TTS, more control, easier to debug). Key insight: latency is the constraint. Humans expect responses in 500ms. Every millisecond matters. No market adoption statistic or universal latency benchmark is established by this skill. Detailed Guide Read [the detailed guide](references/detailed guide.md) before executing this skill. It retains the complete procedure and reference material. Treat its safety, prerequisites, and validation requirements as mandatory. For focused work, load the relevant sections; for end to end work, read the guide completely. Legacy Pipeline Sketch (not a production implementation) Optimization Tips: Start TTS while LLM still generating (streaming) Pre compute first response segment during user speech Use Flash/turbo models for latency Voice Activity Detection Pattern Detect when user starts/stops speaking When to use : All voice agents need VAD for turn taking VOICE ACTIVITY DETECTION (VAD): When to Use User mentions or implies: voice agent User mentions or implies: speech to text User mentions or implies: text to speech User mentions or implies: whisper User mentions or implies: elevenlabs User mentions or implies: deepgram User mentions or implies: realtime api User mentions or implies: voice assistant User mentions or implies: voice ai User mentions or implies: conversational ai User mentions or implies: tts User mentions or implies: stt User mentions or implies: asr Limitations Use this skill only when the task clearly matches the scope described above. Do not treat the output as a substitute for environment specific validation, testing, or expert review. Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing. Version, privacy and implementation boundaries These mixed version SDK sketches are conceptual and were not executed. Deepgram constructor/transcription.live, ElevenLabs stream signatures, the beta Realtime client and the named Silero JavaScript package must be checked against the installed official SDKs before use; do not infer missing package/runtime availability. smart format formats text and is not noise reduction; model keyword/keyterm flags vary. A model VAD probability threshold cannot be set from raw amplitude without a calibrated mapping. Confidence scores/latency figures are not universally comparable. Define end of user turn to first audible output and measure p50/p95 on representative devices/networks/languages. STT first partial, LLM first token and TTS first audio times are not simply additive end to end guarantees. Do not send partial/stale hypotheses to irreversible tools: serialize turns, deduplicate final transcripts, buffer complete TTS safe segments, maintain history, and cancel/discard obsolete generation/audio. On barge in truncate unheard assistant content in server history as well as stopping local playback; a checkpoint alone is insufficient. Clients need appropriate ephemeral/session credentials, not long lived API keys. Obtain authorized microphone/recording/voice/recipient scope, minimize audio/transcript retention and external provider transfers, and avoid voice cloning/impersonation without rights. Bound sockets, retries, queues and costs; clean up listeners/streams on disconnect. Spoken names/numbers and claimed identity are not authentication. Validate tenant/recipient/action and obtain transaction specific confirmation where needed. Text cleanup must preserve decimal numbers, currency, URLs and code semantics; SSML requires supported engines and escaping. A voice prompt is not a tool authorization boundary. No call, recording or provider request was made.

voice-agents