Guide · Pipecat

One processor, two frame tags.

Pipecat pipelines are a list of frame processors, which makes the split straightforward: the processor emits each audio frame twice, tagged raw and enhanced, and downstream services declare which tag they want.

Install

pip

pip install anecho-pipecat

Pipeline · telephony

python · pipecat

import os
from pipecat.pipeline.pipeline import Pipeline
from anecho_pipecat import AnechoProcessor

anecho = AnechoProcessor(
    api_key=os.environ["ANECHO_API_KEY"],
    model="clearline-s",
    sample_rate=8000,           # the call is narrowband; keep it that way
    echo_suppression=True,
    split_pipeline=True,
)

pipeline = Pipeline([
    transport.input(),
    anecho,                     # emits raw + enhanced, sample-aligned
    stt,                        # consumes tag="raw"
    context_aggregator.user(),
    llm,
    tts,
    transport.output(),
])

Turn-taking on the clean channel

Point your VAD analyser at the enhanced tag. This is the half of the split our own measurements support most strongly: voice-activity F1 barely moves, but the false-alarm rate drops by twenty to thirty points in babble and competing-speaker conditions — which is exactly the failure where someone else in the room stops your agent mid-sentence.

python · VAD source

from pipecat.audio.vad.silero import SileroVADAnalyzer
from anecho_pipecat import AnechoVADSource

transport = DailyTransport(
    ...,
    params=DailyParams(
        audio_in_enabled=True,
        vad_analyzer=SileroVADAnalyzer(),
        # take VAD input from the enhanced tag rather than the wire
        vad_audio_source=AnechoVADSource(anecho, channel="enhanced"),
    ),
)

Quality scoring as a pipeline event

Nyquist runs alongside and emits a frame you can log or act on. The useful pattern is not a dashboard — it is a runtime decision: when the score drops below your threshold, stop trusting the transcript and ask the caller to repeat, before the language model acts on a hallucinated account number.

python · react to call quality

@anecho.event_handler("on_quality_frame")
async def on_quality(processor, frame):
    if frame.score < 0.35:
        await task.queue_frame(
            TTSSpeakFrame("Sorry — the line is breaking up. Could you repeat that?")
        )

LiveKit guide API reference Why this is the default