pipecat-siphon

A pipecat frame serializer for the media WebSocket protocol of the SIPhon RTP engine (siphon-rtp).

The SIPhon RTP engine is a media engine that can bridge a call leg's audio to an external WebSocket server: it decodes RTP to linear PCM, streams it up, and encodes PCM coming back down into RTP toward the caller. Your bot never touches RTP, jitter buffers, or codecs. This package is the small adapter that lets a pipecat pipeline read and write that wire.

The SIPhon RTP engine keeps its own native wire on purpose. Every telephony vendor has a different one, and serializers exist precisely so an engine does not have to pretend to be somebody else. There is no Twilio/Telnyx emulation mode here, and there will not be one. The adapter lives on this side, and it is small: the engine's binary-L16-plus-JSON-envelope wire is simpler than the base64-in-JSON formats pipecat already ships support for.

Pipecat's contributing guide directs new service and transport integrations to community-maintained external packages rather than pull requests into the core repository, so this is the sanctioned shape for an integration like this one.

Install

pip install pipecat-siphon

It depends on pipecat-ai>=1.8.0,<2 and nothing else. Python 3.11 or newer, matching pipecat.

The shape of the integration

The engine dials out. The SIPhon RTP engine is the WebSocket client; your bot is the server. So the example under examples/ stands up a server and waits, rather than connecting anywhere.

from pipecat.pipeline.pipeline import Pipeline
from pipecat.transports.websocket.server import (
    SingleClientWebsocketServerParams,
    SingleClientWebsocketServerTransport,
)
from pipecat_siphon import SiphonFrameSerializer

transport = SingleClientWebsocketServerTransport(
    host="127.0.0.1",
    port=9001,
    params=SingleClientWebsocketServerParams(
        audio_in_enabled=True,
        audio_out_enabled=True,
        serializer=SiphonFrameSerializer(),
    ),
)

pipeline = Pipeline([transport.input(), your_processors, transport.output()])

That transport holds one connection for the life of the process, which is the right shape for examples/echo_bot.py and the wrong one for a phone number: the engine dials a fresh WebSocket per call, so a single-client server is a hard ceiling of one call at a time. For anything that answers a real number, serve a WebSocket per call and build the pipeline inside the handler -- examples/agent_bot.py does exactly that with pipecat's FastAPIWebsocketTransport. The serializer is per call either way; it carries the call's own start parameters.

Then point the engine at it, in the profile of a native-JSON offer:

{
  "id": 1,
  "command": "offer",
  "call_id": "call-7@198.51.100.2",
  "from_tag": "caller",
  "sdp": "v=0\r\n...",
  "profile": { "ws_uri": "ws://127.0.0.1:9001/stream", "ws_vad": true, "ws_barge_in": true }
}

ws_uri is takeover: the WebSocket server is the far side of the call. ws_tee (or the attach_ws_tee verb) is a tee: the call keeps relaying between two real parties and a copy of the decoded audio is streamed out, send-only. Both use the same wire, and this serializer handles both; on a tee it refuses to write audio back, because the engine would not inject it anyway.

Which callers can reach a bot

A takeover makes the engine the caller's only peer, so the engine has to terminate whatever the caller negotiated and say so in the answer it writes back. answer_local writes that answer itself; offer/answer rewrites it from the B leg's SDP, and a takeover call has no B leg. So the verb decides what the caller may be (SIPhon RTP engine 0.3.0):

On a supported secure takeover the engine mints its own keying rather than echoing the caller's, decrypts the caller's ingress before it becomes the audio your bot hears, and encrypts the downlink on the way back out. It is fail-closed: nothing leaves in the clear toward a peer that negotiated encryption. Every refusal comes back at offer time with a stable token at the front of the reason, so a controller can branch on it without parsing prose. Before 0.3.0 those same offers returned ok and produced a call that answered and bridged nowhere, which is worth knowing if you are moving a bot off an older engine.

None of this reaches the WebSocket. The bot sees the same start envelope and the same L16 frames whether the caller is a plaintext SIP phone or a WebRTC endpoint.

Attaching a bot to a call that is already up

ws_uri puts a bot on the call at negotiation time and keeps it there until the call ends. From SIPhon RTP engine 0.4.0 a controller can also do it mid-call, with two verbs:

• attach_ws_bridge (call_id, from_tag, ws_uri) on a call that has no bridge takes a live two-party relay over: leg A's audio goes to your bot and the A↔B path is unwired, so the other party hears nothing until it is detached. On a call that already has one it re-points it at a different server without renegotiating — codec, wire rate, VAD, echo cancellation, source gate, SRTP keying and ICE selection all carry across, so there is no re-INVITE and nothing reverts to a default.
• detach_ws_bridge puts the two parties back together, reinstalling the exact forward rules the takeover displaced.

For a pipecat deployment that is the difference between "the bot answers the call" and "the bot joins a call in progress" — an agent taking over from a human, a supervisor handing a caller to a bot, or a bot swapped out mid-conversation. From this side of the socket nothing changes: your server gets the same start envelope and the same audio, and this serializer does not care which verb put it there.

The same release added ws_bridge_started / ws_bridge_ended events. The end reason (detached, server_closed, server_stopped, call_ended, transport_error) is how a controller finds out the bot's socket died — worth wiring up, because a takeover bridge is the caller's only far side, so a bot that goes away is a live call with nobody on it.

The wire

Two frame kinds share one socket:

• binary frames carry raw audio in the negotiated format, one frame per ptime, no base64;
• text frames carry a {"type": "<snake_case tag>", "data": {...}} control envelope whose data object uses camelCase field names.

The first text frame is always start:

{
  "type": "start",
  "data": {
    "streamId": "ws-call-7@198.51.100.2",
    "callId": "call-7@198.51.100.2",
    "direction": "duplex",
    "media": {
      "encoding": "L16", "sampleRate": 8000, "channels": 1,
      "bitDepth": 16, "endianness": "little", "ptime": 20
    }
  }
}

Sample rates: the wire rate is negotiated, and authoritative

media.sampleRate in the start envelope is the wire rate, in both directions, for the life of the stream, and it is authoritative. Never assume 8000. Frame against the number start gives you, and expect it to differ from what your pipeline runs at.

From SIPhon RTP engine 0.3.0 that rate is selectable independently of the call's codec, so an 8 kHz G.711 caller can stream 16 kHz L16 to a bot that wants wideband input:

Valid rates are multiples of 1000 from 8000 to 48000. Anything else fails the offer, answer or attach with a typed error before the call is dialled — the engine never silently clamps to a rate you did not ask for. Leave the knob out and the wire follows the leg's codec rate (8000 for G.711, 16000 for G.722/AMR-WB) with no conversion built at all, exactly as before 0.3.0.

Set it to whatever your pipeline runs at. The conversion then happens in the engine instead of here, which is the better place for it: the SIPhon RTP engine resamples frame by frame at the RTP boundary with nothing held back, while pipecat's SOXR stream resampler at VHQ swallows the first few chunks and then delivers in bursts. Same audio, steadier timing, and one less stage between the caller and the model.

On 0.2.1 and earlier the knobs do not exist. An older engine ignores the field rather than rejecting it, so you get the codec rate back with nothing to point at — check the engine version, or just read the rate out of start, which is right on every build.

This serializer handles the mismatch for you. setup(StartFrame) records the pipeline's audio_in_sample_rate; the wire rate arrives later, in start (the engine cannot dial you before you are listening, so the ordering is always pipeline-first). Every frame is then resampled between the two with pipecat's own create_stream_resampler, in both directions, and a media_renegotiate mid-stream moves the wire rate without restarting anything. When the rates already match, nothing is resampled and the bytes pass through untouched.

If you want to fix the wire rate before start arrives, set wire_sample_rate in the params; start still wins the moment it lands.

Stereo and track separation

A tee with channels: 2 interleaves the two legs: channel 0 is the caller, channel 1 is the callee. The stereo_input param decides what the pipeline sees:

Outbound, if the wire is stereo the bot's mono audio is written to both channels: a bot has one voice, and silencing a channel would be a stranger default than duplicating it.

Message matrix

Engine → pipecat (deserialize)

Pipecat → engine (serialize)

Deliberately unmapped

• A–D DTMF digits. RFC 4733 defines events 12–15 as the A–D digits and the engine will report them, but pipecat's KeypadEntry only models 0–9, * and #. Nothing downstream could consume an invented frame, so those digits are logged at debug and dropped.
• play_start / play_stop. Downlink audio needs no announcement on this wire: binary frames are enough, and the engine's v1 takeover bridge rejects an inline (base64) play_start with an error anyway. There is no pipecat frame that means "start a named playback segment", so nothing is emitted and nothing is consumed.
• event inbound. An opaque application payload has no pipecat frame it corresponds to; it is logged rather than guessed at.
• FrameSerializer.type / FrameSerializerType. Older pipecat releases had a type property on serializers. It does not exist in pipecat 1.8 — the transport decides binary versus text from whether serialize returned bytes or str — so this package does not define one.

Turn taking and barge-in

Engine profile flags move the turn work into the media engine, which is usually where you want it for telephony:

With ws_vad on and the default speech_frames="vad", the engine's VAD drives pipecat's turn machinery directly and you do not need a local VAD in the pipeline at all — no SileroVADAnalyzer on the transport, no second copy of the audio being classified.

Which detector: the energy gate answers "is something loud here", so as a turn detector it fires on mains hum, breathing and fan noise. "neural" runs a Silero v5 forward pass inside the engine and answers "is what is here speech" instead, for about 27 µs per 20 ms frame on an 8 kHz leg, with a 32 ms detection floor from the network's own window. For a conversational bot, "neural" plus ws_vad_min_speech_ms somewhere in 60–120 ms is the pairing that stops a cough, a door or one burst of echo cutting the bot off mid-sentence. A detector the engine cannot build for the leg fails the offer rather than quietly falling back to the one you were avoiding.

echo_cancellation is not optional once barge-in is on. Echo of speech is speech to any detector, the neural one included, so on a handsfree or loudspeaker endpoint the bot's own voice returning up the caller's uplink reads as the caller interrupting.

If you are pinned to SIPhon RTP engine 0.3.0, upgrade. That release counted the energy detector's trailing hangover in milliseconds rather than in ptime frames, so speech_stopped — and the VADUserStoppedSpeakingFrame this serializer emits from it — landed twenty times too late at a 20 ms ptime, i.e. never inside a normal turn, while speech_started and barge-in kept working and made turn taking look healthy. Fixed in 0.3.1.

Put together, the profile for a conversational bot on 0.3.0 — wideband wire on a narrowband call, engine-side turn taking, no local VAD:

"profile": {
  "ws_uri": "ws://198.51.100.10:9001/stream",
  "ws_sample_rate": 16000,
  "ws_vad": true,
  "ws_barge_in": true,
  "ws_vad_engine": "neural",
  "ws_vad_min_speech_ms": 100,
  "echo_cancellation": true
}

If you would rather drive the turn explicitly, set speech_frames="user" and wire ExternalUserTurnStartStrategy / ExternalUserTurnStopStrategy. That strategy is constructed with interruptions disabled, so auto_interrupt=True (the default) also emits an InterruptionWorkerFrame after the UserStartedSpeakingFrame. Pipecat's serializer API hands the transport one frame per WebSocket message, so the second frame is queued and drains on the next message — at most one ptime later, since the engine's takeover ticker never stops sending.

Barge-in in the other direction is automatic: any InterruptionFrame in the pipeline becomes a clear on the wire.

Params

All of these live on SiphonFrameSerializer.InputParams, which extends pipecat's own FrameSerializer.InputParams (so ignore_rtvi_messages and resampler_clear_after_secs are there too).

Example

examples/echo_bot.py is the minimum that proves the path: a WebSocket server, this serializer, and a processor that echoes the caller's audio back into the call. Run it and point an engine at it:

python examples/echo_bot.py --host 127.0.0.1 --port 9001

Then offer a call with "profile": {"ws_uri": "ws://127.0.0.1:9001/stream"} and you hear yourself back, one jitter-buffer frame plus one playout frame later.

examples/agent_bot.py is that same path with the echo processor replaced by an STT → LLM → TTS chain, so a caller talks to a model instead of to themselves. It needs three API keys in the environment and pip install "pipecat-ai[anthropic,deepgram,cartesia]"; the vendors are one line each to swap, and the file explains which knobs are there for latency rather than for taste. The turn taking is the engine's, not the pipeline's — there is no VAD analyzer in it at all.

Given --control-url, it also connects to the SIPhon SIP engine's control plane and gets an end_call tool, so the model can hang up when the conversation is done rather than emitting a magic phrase for something else to notice. With BOT_TRANSFER_TARGET set it gets a transfer_call tool as well, which REFERs the caller to that destination and keeps the call until the transfer's verdict: many phones refuse a REFER from the network, and then the bot stays on the line, is told the transfer did not work and tells the caller, rather than leaving them in silence (a phone that never answers counts as refusing after BOT_TRANSFER_VERDICT_TIMEOUT, 60 s by default). With BOT_TRANSFER_MODE=bridge the caller's phone is left out of it: SIPhon rings the destination itself (dial with on_answer: "bridge"), the caller stays with the bot while it rings, and the two are joined when it answers, which drops the bot's stream. A refusal or no answer within BOT_TRANSFER_RING_TIMEOUT (30 s) keeps the caller with the bot, as above. The routing script that hands the call over says where that leg goes and whom it presents, as the call's vars transfer_next_hop, transfer_from, transfer_p_asserted_identity and transfer_privacy.

With --call <sip-uri> (and --call-ws-uri, the bot's own address as the engine reaches it) the bot places one call through the control plane's originate instead of waiting for one, opens it as the caller once it is answered (--call-context says what the call is about), and exits when it is over: 0 for an answered call that ended normally, 1 with the outcome logged for a refused, rejected (rejected 486 Busy Here), unanswered or failed one. The pieces are in the package for a bot of your own: pipecat_siphon.PlacedCalls places a call, adds placed=<channel> to the ws_uri so the engine's dial-in names its call before start arrives, and reports how the call went from the control events, hanging up a call siphon never reports on or one answered into a bot that was never reached. It needs siphon-control 0.9.0 or newer, the first whose originate returns the placed Call. The example's docstring covers the platform side: the media profile the engine's built-in voice_ai does not fully supply, and the correlation detail that decides whether the control channel ever finds its call — the engine expands {call_id} in a ws_uri to the SIP Call-ID, so the two channels join on sip_call_id, not on the control frame's own call_id.

Speaking another language

Left alone, the agent's voice speaks its provider's default language, which is English: a model answering in Dutch is read out by an English voice, with an English accent. Four variables fix that, and with none of them set the bot is built exactly as before:

BOT_TTS_LANGUAGE is the language the call starts in (the greeting included), and goes to every synthesizer as its language setting. With more than one language in BOT_TTS_LANGUAGES the model gets a set_speaking_language tool, limited to that list, and calls it when the caller asks for another language: the voice, its language and a recognizer pinned to one language all change for the rest of the call. BOT_TTS_FOLLOW_CALLER=true also moves the voice to the language the recognizer hears, once two stretches of the caller's speech in a row are in it; that needs a recognizer that reports a language per transcript, which is Deepgram on multi, its default here.

The language setting alone does not remove the accent; the voice has to speak that language natively. A voice reads every language in its own accent. On Cartesia, where a voice is a voice id on your account, give each language a native voice in BOT_TTS_VOICES, or use one of Cartesia's multilingual voices with the language set. Google's voice names carry their locale, so a language without a voice in the map gets the configured voice in that locale (en-US-Chirp3-HD-Charon becomes nl-NL-Chirp3-HD-Charon). OpenAI's speech endpoint takes no language, so its gpt- voices are instead instructed to use a native accent. The local voice, Kokoro, has no Dutch at all, and a language it cannot speak, or a Kokoro voice of another language, stops the bot at startup rather than speaking with an accent.

A bot that answers in Dutch with a native voice and switches to English when asked:

BOT_TTS_LANGUAGES=nl,en
BOT_TTS_VOICES=nl=<a Dutch voice id>,en=<an English voice id>

Looking things up for the caller

Two more tools, each off until it is configured, and the prompt mentions only the ones that are on:

caller_lookup takes no arguments. It fetches BOT_CALLER_LOOKUP_URL with {caller} replaced by the number in the media socket's caller query parameter (6 to 15 digits, no plus; anything else counts as no number), and hands the model the JSON object it gets back: strings, numbers and booleans, and lists or objects of those, one level deep. Nothing else is interpreted, so the description is how the model learns what the record is for. A model that could name the number could be talked into looking up anyone, so it cannot. No caller on the URL, a 404 or any other failure status, an answer that is not a JSON object or is over 2 KB, or no answer within 1.5 s all reach the model as "not available", and the bot says it cannot see that right now. A URL without {caller} stops the bot at startup. The number is logged masked. The example does not verify the URL, so the bot's port must be reachable only by whatever dials it.

weather takes a place the model chooses, a capital when the caller names a country and no town, and reads the current temperature, conditions and wind from open-meteo, which needs no key. It gives up after 3 s.

Calling the caller back

With a control plane, the model can arrange a callback: call_back(minutes, reason) rings the caller back once this call is over, minutes from now, and opens that call as the caller would, about reason.

{number} is the number the network asserted for the caller: the first P-Asserted-Identity on the INVITE siphon handed over that names a number, + and all. The model never names it, and the media socket's caller parameter is not used, because whoever dials the bot can set it and a callback costs money. A call with no asserted number, or one the caller withheld (Privacy: id), is told the number is not available. There is one callback per number at a time. It needs --control-url and BOT_CALL_WS_URI, and starting with BOT_CALL_BACK_URI but without either stops the bot. Callbacks are held by the running bot, so stopping it loses any not yet placed, and each one lost is logged as an error. Numbers are logged masked.

One bot, many configurations

BOT_CONFIGURATIONS_DIR points at a directory of <key>.yaml files. A call that names a key is served with that file's settings over the bot's own:

# CUST-0042.yaml
system_prompt: You answer the phone for a dental practice.  # or system_prompt_file, beside it
transfer_target: sip:front-desk@example.invalid
caller_lookup_url: https://accounts.example.invalid/patients/{caller}
caller_lookup_description: Look up the caller's next appointment.

The key comes from whoever handed the call over, never from the media socket's URL. Behind siphon-pbx it is the organisation's external_ref from the brief's context (the bot's context setting must be organisation or full). Behind siphon-sip it is the bot_configuration var the routing script hands the call over with. siphon-sip dials the stream before it hands the call over, so the bot accepts the stream and then waits up to 3 s for the handover. It only waits when this directory is set.

The same sources give the caller's number to caller_lookup, in place of the URL's caller: context.caller from siphon-pbx, the handover INVITE's P-Asserted-Identity from siphon-sip. A withheld caller gives none. When siphon-pbx names the organisation, the model is told its name and the local time in its time zone. Nothing the caller sets, a display name included, reaches the prompt.

A call that names no key, a key with no file, or a file that is wrong is served with the bot's own settings; the last is logged as an error. Every file is checked at startup, and one that is wrong stops the bot. Files are re-read per call, so an edit needs no restart. Languages and voices stay the bot's own.

Behind siphon-pbx

When a siphon-pbx flow's voice block hands a call to the bot, the engine dials it with &session=<token> on the URL. With both of these set, the bot fetches that session's brief before it speaks, and reports how the conversation went back to the PBX, so the flow leaves the block on the report instead of waiting out its deadline:

Set both or neither; one without the other stops the bot at startup. The brief's instructions are added to the system prompt for that call. When the block declares options, the model gets a choose_option tool limited to them, which reports the caller's choice once the line passing them on has played. end_call reports the conversation finished instead of hanging up, because what happens to the call next is the flow's decision. A pipeline error reports the bot failed. Each call reports once. After a report the PBX took, the bot waits up to 10 s for the PBX to close its stream before ending the pipeline itself. After one it did not take, the bot ends the pipeline at once. A brief that is refused, unreachable within 2 s, or not the documented shape is logged, and the call is answered with the bot's own configuration. Neither the secret nor the token is logged. pipecat_siphon.VoiceSessionClient and VoiceSession are the same pieces for a bot of your own.

examples/probe.py speaks the engine's half of the media wire — connect, start, stream silence — and reports how much audio comes back and how long the first frame took. No phone, no SIP stack, no engine. It is the fastest way to tell "the AI is broken" from "the media path is broken", because every failure above the media path looks identical from outside: a socket that connects, a bot that reports healthy, and silence.

python examples/probe.py ws://127.0.0.1:9001/stream

Quickstart: a real phone number

examples/quickstart/ is the whole thing end to end: the SIPhon SIP engine (siphon-sip) registers to a SIP provider, inbound calls on that registration are handed to the bot, and the bot can hang up. Nothing in it is provider-specific — registration is RFC 3261, and every credential comes from the environment.

It runs as three containers, so a clone, a .env and one command is the whole setup:

cd examples/quickstart
cp .env.example .env      # three API keys, the SIP account, the address the provider can reach
docker compose up --build

There is a by-hand path for the three processes too, and a probe that tells you whether the bot answers with audio before you spend a phone call finding out. See examples/quickstart/README.md.

Integration harness

integration/ holds a four-way harness that runs the whole chain with nothing mocked: SIPp places a call through the SIPhon SIP engine (siphon-sip), the SIPhon RTP engine (siphon-rtp) bridges the leg to a pipecat bot using this serializer, and the assertions are on the audio that comes back rather than on the socket being open. Three scenarios: a known signal survives the round trip, checked spectrally against a tshark capture; the engine's VAD edges land where the fixture puts them and barge-in flushes the bot's queued speech; and the same call again with ws_sample_rate: 16000, where the wire runs at 16 kHz over an 8 kHz G.711 leg and the audio has to come back at the pitch it went out at.

cd integration && ./run.sh

It needs Docker and nothing else: both SIPhon engines are pulled as their published release images, so this runs on a fresh clone. Point SIPHON_RTP_PATH / SIPHON_SIP_PATH at a checkout to build either from source instead. Not part of CI yet; see integration/README.md.

Compatibility

The control-frame fixtures under tests/wire_fixtures.py are byte-exact strings produced by compiling the engine's own Rust definition against serde_json — not by this package's encoder. A shared bug on both sides of an encode/decode pair sails straight through a round-trip test; it does not survive a comparison against bytes another implementation produced.

Note on version skew: an unrecognised control type is dropped with a warning rather than raising, so a bot pinned to this release keeps working against a newer engine that adds a message.

Development

uv venv --python 3.13 .venv
uv pip install -e ".[dev]"
.venv/bin/python -m pytest
.venv/bin/ruff check .
.venv/bin/ruff format --check .
.venv/bin/mypy

Maintainer

pipecat-siphon is maintained by SIPhon's maintainer at Real Time Telecom B.V., the founding sponsor of the SIPhon project. Commercial support is available through RTT.

License

MIT. See LICENSE.