Silence-Based EOS - rapida.ai documentation

Silence-Based EOS triggers end-of-speech after a fixed duration of silence following the last speech activity or transcript. No ML model, no inference overhead. Provider identifier: silence_based_eos

Source Location

api/assistant-api/internal/end_of_speech/internal/silence_based/
└── silence_based_end_of_speech.go

How It Works

A single worker goroutine manages a timer
When a final SpeechToTextPacket arrives, text is accumulated and the timer is set to timeout
When an interim SpeechToTextPacket arrives, the timer resets
When VAD is active, the timer stores a pending final instead of completing immediately
When VAD ends, pending EOS fires immediately; if VAD end is missing, the recovery timer fires it
State resets for the next turn

STT final → accumulate text → start timer (1000ms)
STT interim → reset timer
VAD start → block completion
VAD end → flush pending final
...silence...
Timer fires → EndOfSpeechPacket → LLM starts

Parameters

Option Key	Default	Range	Description
`microphone.eos.timeout`	`1000 ms`	500–4000 ms	Silence duration before triggering end-of-speech

The UI sets a default of 700 ms when switching to this provider. The backend default (when no value is configured) is 1000 ms.

Setup

No model files or downloads required. Works out of the box with both Docker and source builds. To select this provider, set microphone.eos.provider to silence_based_eos in the assistant’s audio options (or leave it unset — it is the default).