Give your voice agent the context to wait, check back or prepare a response. Turn live speech into clear, revisable turn-taking advice.
A ten-second turn-taking walkthrough. At two seconds, “Let me think” is followed by silence and the turn stays protected. At four seconds, the answer continues. After “That's all”, a turn-end candidate is proposed at six seconds. Speech resumes at eight seconds, revoking the candidate. Waveform peaks show speech; the signal flattens during silence. Advice remains revisable, and the application controls the response.
minutes of real conversation
used to train our models
Listen to the signalSpeech, pauses and endpoint evidence
Make room for contextOptional transcript and answer cues
Keep your app in controlStructured advice, clear reasons
Explore how the advice changes with a thinking pause, a complete answer or a speaker picking up their thought.
“Let me think about that…”
Listen to the speaker
While speech is observed, the policy holds the turn open. A pause has not become a decision.
Each decision has a reason and a state. This walkthrough shows selected fields from the EOT advice contract; your application retains control of playback.
{
"schema": "sharedpro.eot.advice.v1",
"type": "eot_advice",
"action": "WAIT",
"event_type": "state_change",
"state": "HOLD",
"reason_codes": [
"speech_observed"
],
"mode": "shadow",
"stage": 3,
"advisory": true,
"actionable": false
}Combine the sound of a pause with the context around it. Keep the decision open when the evidence changes.
Requests for thinking time, unfinished phrases and self-corrections can protect an answer. Add caller-supplied transcript or context cues alongside the audio.
Use observed continuation pauses within a session to extend waiting. Preserve time already granted to the speaker.
Speech resumes. Context changes. A candidate can be revoked. Your controller receives the update and decides how to respond.
Inspect reason codes, source-aligned timing, evidence availability and candidate identity. Replay microphone or file sessions in the dedicated lab and export the decision trace.
Connect a dedicated turn-taking service alongside your existing speech and language models. Your application owns the conversation.
The authenticated WebSocket endpoint accepts mono 16 kHz PCM. The Python client handles framing and event reads; your backend keeps credentials and conversation control.
/api/eot/liveAdvice stays advisory. Speech-to-text, response generation, playback and interruption handling belong to your application.
{
"type": "start",
"schema": "sharedpro.eot.stream.v1",
"encoding": "pcm_f32le",
"sample_rate": 16000,
"channels": 1,
"mode": "shadow",
"stage": 0,
"answer_id": "answer-1",
"answer_shape": "open",
"capture_epoch": 0,
"clock_domain": "capture_samples",
"research_capture": false
}Begin in observation mode. After ready, send EOT2-framed participant PCM and read events concurrently. Enable further advisory stages through your serving configuration.
Bring turn advice into the voice experiences you’re building.
Account for open answers, short responses, requests for more time and a speaker correcting themselves.
Give a learner time to form an answer and keep the feedback moment in your application’s control.
Combine acoustic evidence and explicit conversation cues when deciding the next conversational move.
Built as a focused, inspectable service for your voice stack.
The API combines speech activity and pause timing with supported pitch and energy changes, endpoint-model evidence, answer type and optional transcript or context cues. Evidence and policy advice are returned separately, so your application can inspect the basis of a recommendation.
It is a standalone authenticated WebSocket API with a Python client. Feed it participant audio alongside your existing transcription path, and optionally provide transcript revisions or context. Your speech-to-text, language model and text-to-speech services stay in your stack; this API supplies turn advice.
Explicit cues such as a request to think or an unfinished phrase protect the turn. Observed continuation pauses within the same session can extend waiting. That history does not shorten a previously granted extension or create a speaker profile across sessions.
A COMMIT event proposes a possible handover; it is still advisory. If speech resumes, context changes or source evidence becomes invalid, the API can revoke the candidate. Your application owns playback, interruption handling and cancellation.
Use the decision walkthrough on this page to explore the behavior. The dedicated project lab supports microphone input, file replay and exportable decision traces. Contact us to discuss the developer preview and evaluate the stages relevant to your use case.
Our models were trained on 300,000 minutes of real conversation. The integration provides shadow observation, replay and decision traces to evaluate behavior in your own setting. The current API is an experimental advisory preview; production qualification and application behavior are separate from training scale.
Explore the Turn-Taking Decision API for your voice application.