Pipecat

One person, near the microphone

Focus keeps one voice: the closest one, or the voice of a print. It is made for a caller on a phone, a headset or a laptop, not for a microphone in the middle of a room.

pipecat-alkove puts Alkove in a Pipecat bot: AlkoveFilter keeps the user’s voice (Focus), AlkoveVADAnalyzer finds where they speak (Cue). Both run on the bot’s CPU, with no server and no key.

Install

pip install pipecat-alkove

The models (Focus 1.1 MB, Cue 1.1 MB) are downloaded once into the cache, checked by their sha256: the analyzer loads them when it is created, as Silero’s does, the filter when the transport starts. Pass models=alkove.Models.load(["focus", "cue"], dir=...) to read them from your own folder.

Use

from pipecat.transports.base_transport import TransportParams
from pipecat_alkove import AlkoveFilter, AlkoveVADAnalyzer

params = TransportParams(audio_in_enabled=True, audio_in_filter=AlkoveFilter())
vad = AlkoveVADAnalyzer()  # VADParams(confidence=0.5) unless you pass params

Give the analyzer to your user aggregator (LLMUserAggregatorParams(vad_analyzer=vad)) or to a VADProcessor, as you would Silero’s.

AlkoveFilter

• It keeps the closest voice; with voiceprint= (an alkove.Voiceprint or its 192 numbers), the voice of that print. strength= and normalize= work as in the web library. It works at the transport’s input rate.
• FilterEnableFrame(False) gives the raw microphone back at the same delay; FilterUpdateSettingsFrame(settings={"strength": 0.5}) and {"voiceprint": ...} change it while it runs.
• It returns the samples that are ready: a little less than the input at the start (Focus’s 16 ms), then about the input’s size. The chain is the one the web library’s inputs run.

To keep one person’s voice rather than the closest one, make their print with Timbre, from 1.5 s of their voice or more at 16 kHz:

from alkove import Models, Timbre

models = Models.load(["focus", "cue", "timbre"])
me = Timbre(models).embed(pcm)
f = AlkoveFilter(models=models, voiceprint=me)  # an alkove.Voiceprint or its 192 numbers

AlkoveVADAnalyzer

• Cue’s speech probability every 30 ms, at any input rate.
• Where speech starts and ends is decided by Pipecat, from VADParams. Cue was tuned with a confidence of 0.5, so that is its default.
• stop_secs stays Pipecat’s 0.2 s: its turn strategy (Smart Turn) decides when the turn ends, and Pipecat’s latency figures assume 0.2.
• Behind AlkoveFilter, it hears only the user.

example.py runs both on your microphone and prints when you speak. On macOS, pyaudio needs brew install portaudio first.

Tested with pipecat-ai 1.13.0. The plugin is maintained by Alkove, not by the Pipecat team: source.