Higgs Realtime in Pipecat: three services become one

We’ve shipped pipecat-boson, which brings Higgs Realtime into Pipecat as a Realtime LLM option. It’s on PyPI now:

$ uv add pipecat-boson

A typical Pipecat voice agent is three services chained together. Speech-to-text turns audio into a transcript. A language model turns the transcript into a reply. Text-to-speech turns the reply into audio. Each one is a separate vendor, a separate network hop, and a separate place where latency accumulates and things break.

It also means everything that isn’t a word gets thrown away at step one. Hesitation, frustration, urgency, the difference between “…yes” and “yes!”—transcription discards all of it before your model ever sees it.

BosonRealtimeLLMService replaces all three. It takes live audio in and streams audio back, handling understanding, reasoning, and speech generation inside a single model. In your pipeline, it goes where your STT, LLM, and TTS services used to be:

import osfrom pipecat_boson.realtime import BosonRealtimeLLMService llm = BosonRealtimeLLMService(    url=os.environ["BOSON_REALTIME_URL"],    api_key=os.environ["BOSON_API_KEY"],    model="higgs-realtime",    voice="default",    instructions="You are a concise and helpful voice assistant.",)

Then drop it into the pipeline:

pipeline = Pipeline([    transport.input(),    user_aggregator,    llm,    transport.output(),    assistant_aggregator,])

That’s the whole integration.

The obvious win is operational: one vendor instead of three, one bill, one thing to debug. The less obvious win is that it’s what makes fast interruption possible at all.

When a user cuts in mid-sentence on a chained pipeline, you have to detect the interruption, kill the TTS playback, discard the audio that was queued but never heard, and reconcile your conversation state against what the user actually heard versus what you sent. Every one of those steps costs time and is a place to get it wrong.

No platform fee, no per-seat pricing, and nothing extra for using it through Pipecat. New accounts come with free credit, so you can build something before deciding whether you like it.

Worth saying why that number matters beyond the line item: at $15–20 an hour, real-time voice only makes sense replacing expensive human labor. At under a dollar, you stop designing around minimizing talk time. Agents that stay in the conversation, practice and role-play tools that need many hours of low-stakes dialogue, per-person rather than per-company deployments—those become viable at a price point they weren’t before.