Pipecat bridge

contrib · v0.1.0

Run a Pipecat pipeline behind a chanx transcriber topic, so any of Pipecat's speech-to-text services speaks the transcriber contract. chanx keeps the socket; Pipecat never sees a transport.

Install	copit add @chanx-kit/pipecat-bridge
Import from	pipecat_bridge
Requires kits	audio-stream-in
Contract	transcriber@1
UI kits	transcriber
Python packages	chanx>=2.10.2,<3, pipecat-ai>=1.3,<2
Keywords	audio, speech-to-text, pipecat, bridge
Source	kits/pipecat_bridge

Run a Pipecat pipeline behind a chanx transcriber topic. Pipecat integrates dozens of speech services; this bridge makes any of its speech-to-text services speak the transcriber contract, so the transcriber and voice-agent UI kits work with it unchanged.

copit add @chanx-kit/pipecat-bridge

from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.processors.audio.vad_processor import VADProcessor
from pipecat.services.deepgram.stt import DeepgramSTTService   # pipecat-ai[deepgram]

from .ws_kits.audio_stream_in import AudioConfig
from .ws_kits.pipecat_bridge import PipecatTranscriberTopic


class MicTopic(PipecatTranscriberTopic):
    # Silero takes 8 or 16 kHz; the browser captures at whatever audio_config says.
    audio_config = AudioConfig(sample_rate=16000)

    def processors(self, audio):
        return [
            VADProcessor(vad_analyzer=SileroVADAnalyzer(sample_rate=audio.sample_rate)),
            DeepgramSTTService(api_key=os.environ["DEEPGRAM_API_KEY"], sample_rate=audio.sample_rate),
        ]

The VADProcessor is what gives turns: its speaking frames become speech_started and utterance_end, which the voice agent needs for turns and barge-in. Without it you get partials and finals, and one utterance per audio_start to audio_end.

How it fits

chanx owns the WebSocket, the topic, authentication and limits; Pipecat never sees a transport. Each audio_start builds a pipeline [*processors(audio), sink]: audio chunks are queued in as InputAudioRawFrame, and the sink turns InterimTranscriptionFrame, TranscriptionFrame and the user speaking frames (a transport's, or a VADProcessor's) into transcript_partial, transcript_final, speech_started and utterance_end. Voice activity notices a pause before the service finalizes, so the utterance end is held until the final arrives. audio_end queues an EndFrame, so the service flushes before the pipeline stops.

Pipecat is a large dependency (a few hundred MB with its audio stack), so it belongs to this kit alone; nothing else in chanx-kit needs it.

Status

Verified on Pipecat 1.12 (2026-09-27): the shared transcriber suite passes through the bridge with a scripted Pipecat STT service; the pipeline hop adds a median of 0.2 ms per chunk (p95 0.3 ms); and live, Silero VAD plus Pipecat's DeepgramSTTService turned two spoken sentences into two full turns (speech started, partials, final, utterance end each). Each stream runs its pipeline on its own WorkerRunner. Run the tests with Pipecat installed; they are skipped without it.