Skip to main content
Community-maintained integration. This service is built and maintained by edbrdi. Pipecat does not test or officially support it. Please report issues and request changes on the source repository. Learn more about community integrations.

​
Overview

AlkoveTurnAnalyzer is a turn analyzer that decides when the user has finished speaking. It reads the end of turn head of Cue, the speech detection model of Alkove, an open-source (Apache-2.0) voice SDK. It needs no transcript, no API key and no GPU: Focus then Cue run locally on the CPU with Alkove’s own inference engine, and the head is read every 0.1 s of each pause, as the audio comes. The same package gives turn.interruption_strategy(), a user turn start strategy: while the bot speaks, the user’s turn starts only once their speech spans min_duration seconds, so a “mm” or a “yeah” lets the bot go on.

Source Repository

Code, example and README of the Pipecat plugin

Model Card

Cue’s weights (Apache-2.0), its end of turn evaluation and limits

​
Installation

This is a community-maintained package distributed separately from pipecat-ai:

​
Prerequisites

No account or API key. The models (Focus 1.1 MB, Cue 1.2 MB) download once into the cache and are checked by their sha256, when the analyzer is created.
  • A VAD analyzer in the pipeline, for example AlkoveVADAnalyzer
  • Languages: evaluated in English, French, Spanish, Italian, Portuguese, German and Dutch

​
Configuration

AlkoveTurnAnalyzer(params=AlkoveTurnParams(...), min_duration=0.7)
​
float
default:"0.7"
Seconds the user’s speech must span, short gaps included, to start their turn while the bot speaks. Used by interruption_strategy().
​
int | None
default:"None"
Input sample rate. None takes the transport’s.
​
alkove.Models | None
default:"None"
Models to use. None loads Focus and Cue once per process, shared by every session.

​
Usage

​
How It Works

  • The turn is complete on the first read of the head over threshold once the pause reaches min_pause, or at max_pause. The head starts over at each turn’s end.
  • The analyzer runs Focus then Cue itself, on the audio the user aggregator gets: behind AlkoveFilter, Focus runs twice.
  • interruption_strategy() shares them. While the bot speaks, a shorter “mm” or “yeah” starts no turn, and its transcript does not either. With the bot quiet, or once it stops while the user goes on, the turn starts with the VAD, as Pipecat’s default.

​
Compatibility

Tested with Pipecat 1.15.0. See the changelog for the latest version.

​
Notes

  • The head hears the voice, not the words: it does not know that “my number is” asks for more.
  • Evaluated on LiveKit’s eot-bench, real user turns to voice agents; the results are on the model card.