sipstuff 0.0.3

SIP telephony automation toolkit: place calls via PJSIP, play WAV/TTS audio, record and transcribe responses

sipstuff

---

⚠️ ⚠️ ⚠️ WIP

REWRITE ongoing from https://github.com/vroomfondel/somestuff/tree/main/sipstuff

---

SIP telephony automation toolkit — place phone calls and play WAV files or TTS-generated speech via PJSUA2. Includes speech-to-text transcription of recorded calls via faster-whisper. Supports incoming call handling with auto-answer, live TTS, and real-time transcription.

Overview

Registers with a SIP/PBX server, dials a destination, plays a WAV file (or synthesizes text via piper TTS) on answer, and hangs up. Designed for headless/container operation (uses a null audio device, no sound card required). Supports UDP, TCP, and TLS transports with optional SRTP media encryption. Can also act as a callee — auto-answering incoming calls with WAV playback, real-time TTS, or live transcription.

Prerequisites

PJSIP with Python bindings (pjsua2) must be installed. The main project Dockerfile builds PJSIP from source in a multi-stage build. For local development:

# Debian/Ubuntu prerequisites
sudo apt install build-essential python3-dev swig \
    libasound2-dev libssl-dev libopus-dev wget

# Build and install PJSIP (default: 2.16)
./dist_scripts/install_pjsip.sh

# Or specify a version
PJSIP_VERSION=2.14.1 ./dist_scripts/install_pjsip.sh

Other Python dependencies: pydantic, ruamel.yaml, loguru, tabulate, soundfile, numpy.

Optional extras (installable via pip install sipstuff[extra]):

• cuda — NVIDIA CUDA libraries for GPU-accelerated STT (nvidia-cublas-cu12, nvidia-cudnn-cu12)
• openvino — Intel OpenVINO backend for STT (optimum-intel[openvino])
• training — Voice training utilities (pynput, PySide6)

For TTS support: piper-tts (≥1.4.0) is included as a dependency and works directly with Python 3.14.

For STT support: pip install faster-whisper (optional, only needed for transcription). Whisper models are auto-downloaded on first use (~1.5 GB for the medium model).

For local audio playback: pip install sounddevice (optional, only needed for --play-audio).

CLI Usage

The CLI provides six subcommands: call, tts, stt, callee_autoanswer, callee_realtime-tts, and callee_live-transcribe.

# Via entry point (after pip install)
sipstuff-cli <subcommand> [args...]

# Via module
python -m sipstuff.cli <subcommand> [args...]

tts — Text-to-Speech

Generate a WAV file from text using piper TTS (no SIP server needed). When --play-audio is used without -o, audio is synthesized entirely in memory (no temp file).

# Basic German TTS
python -m sipstuff.cli tts "Hallo Welt" -o hello.wav

# English voice, resampled to 8 kHz for narrowband SIP
python -m sipstuff.cli tts "Hello World" -o hello.wav \
    --model en_US-lessac-high --sample-rate 8000

# Custom model directory
python -m sipstuff.cli tts "Achtung!" -o alert.wav --tts-data-dir /opt/piper-voices

# Generate and play back on speakers
python -m sipstuff.cli tts "Test" -o test.wav --play-audio

# Play without saving to disk (in-memory synthesis)
python -m sipstuff.cli tts "Nur zum Anhören" --play-audio

# Interactive REPL — type text and hear it spoken in real-time
python -m sipstuff.cli tts --interactive --model de_DE-thorsten-high

# CUDA-accelerated TTS
python -m sipstuff.cli tts "Schnell!" -o fast.wav --tts-cuda

Flag	Description
text (positional)	Text to synthesize
--output, -o	Output WAV file path (optional when --play-audio or --interactive is used)
--model, -m	Piper voice model (default: de_DE-thorsten-high)
--sample-rate	Resample to this rate in Hz (0 = native)
--tts-data-dir	Directory for piper voice models (default: ~/.local/share/piper-voices)
--interactive, -i	Interactive REPL: type text and hear it spoken in real-time via speakers (requires sounddevice)
--tts-cuda	Use CUDA GPU acceleration for Piper TTS
--play-audio	Play generated audio on speakers via sounddevice
--audio-device	Sounddevice output device index or name substring
--verbose, -v	Debug logging

stt — Speech-to-Text

Transcribe a WAV file using faster-whisper (no SIP server needed):

# Basic transcription (German, medium model)
python -m sipstuff.cli stt recording.wav

# English, smaller model, JSON output with metadata
python -m sipstuff.cli stt recording.wav --language en --model small --json

# Disable Silero VAD pre-filtering (VAD is on by default)
python -m sipstuff.cli stt recording.wav --no-vad

# CUDA acceleration with large model
python -m sipstuff.cli stt recording.wav --device cuda --model large-v3

# OpenVINO backend
python -m sipstuff.cli stt recording.wav --backend openvino --model base

The --json flag outputs structured JSON including segment timestamps, audio duration, and language probability:

{
  "text": "Ja, mach schon, jetzt ist es auch wurscht.",
  "audio_duration": 17.4,
  "language": "de",
  "language_probability": 1.0,
  "segments": [
    {"start": 0.0, "end": 1.52, "text": "Ja, mach schon,"},
    {"start": 1.52, "end": 3.84, "text": "jetzt ist es auch wurscht."}
  ]
}

Flag	Description
wav (positional)	Path to WAV file to transcribe
--backend	STT backend: faster-whisper (default) or openvino
--model, -m	Whisper model size or HuggingFace model ID (default: medium)
--language, -l	Language code (default: de)
--device	Compute device: cpu or cuda (default: cpu)
--compute-type	Quantization: int8, float16, float32 (auto-selected by default)
--data-dir	Directory for Whisper models (default: ~/.local/share/faster-whisper-models)
--json	Output result as JSON with metadata and segment timestamps
--no-vad	Disable Silero VAD pre-filtering (VAD is on by default, recommended for phone recordings)
--verbose, -v	Debug logging

call — Place a SIP Call

# Minimal — using a config file with a WAV file
python -m sipstuff.cli call --config sip_config.yaml --dest [phone] --wav alert.wav

# Using CLI flags directly (no config file)
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] --wav alert.wav

# Connection via environment variables, only destination and audio via CLI
export SIP_SERVER=pbx.local SIP_USER=1000 SIP_PASSWORD=secret
python -m sipstuff.cli call --dest [phone] --wav alert.wav

# TTS instead of a WAV file (no audio file needed)
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] \
    --text "Achtung! Wasserstand kritisch!" \
    --tts-model de_DE-thorsten-high --tts-sample-rate 8000

# Wait for callee to finish speaking before playback
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] --wav alert.wav \
    --wait-for-silence 1.0

# Record, transcribe, and write JSON call report
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] --wav alert.wav \
    --record /tmp/recording.wav --transcribe \
    --wait-for-silence 1.0

# Record both RX and TX, then mix to stereo
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] --wav alert.wav \
    --record /tmp/rx.wav --record-tx /tmp/tx.wav \
    --mix-mode stereo --mix-output /tmp/mixed.wav

# TTS with a custom model directory
python -m sipstuff.cli call \
    --config sip_config.yaml --dest [phone] \
    --text "Server offline!" \
    --tts-data-dir /opt/piper-voices

# Calling a SIP URI directly (instead of a phone number)
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest sip:[e-mail] --wav announcement.wav

# TLS transport with SRTP encryption and playback options
python -m sipstuff.cli call \
    --config sip_config.yaml \
    --transport tls --srtp mandatory \
    --dest [phone] --wav alert.wav \
    --pre-delay 1.5 --post-delay 2.0 --inter-delay 1.0 --repeat 3 \
    --timeout 30 -v

# Record with explicit STT options
python -m sipstuff.cli call \
    --config sip_config.yaml --dest [phone] --wav alert.wav \
    --record /tmp/recording.wav --transcribe \
    --stt-model small --stt-language en \
    --stt-data-dir /opt/whisper-models

# Live transcription of remote audio during the call
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] --wav alert.wav \
    --live-transcribe -v

# Interactive live TTS — type text during the call that gets spoken in real-time
python -m sipstuff.cli call \
    --server 192.168.1.100 --user 1001 --password secret \
    --dest [phone] \
    --interactive \
    --piper-model /path/to/de_DE-thorsten-high.onnx \
    --text "Hallo, hier spricht die Maschine." \
    --play-audio --play-tx --real-capture -v

# Stream remote audio to a Unix socket (for external processing)
python -m sipstuff.cli call \
    --server pbx.local --user 1000 --password secret \
    --dest [phone] --wav alert.wav \
    --audio-socket /tmp/sip_audio.sock

# NAT traversal — STUN + ICE behind a NAT gateway
python -m sipstuff.cli call \
    --server pbx.example.com --user 1000 --password secret \
    --stun-servers stun.l.google.com:19302,stun1.l.google.com:19302 \
    --ice \
    --dest [phone] --wav alert.wav -v

When --transcribe is used with --record, a JSON call report is written next to the recording (e.g. /tmp/recording.json) and also emitted to stdout/loguru for K8s log collection.

call CLI Flags

SIP connection:

Flag	Description
--config, -c	Path to YAML config file
--server, -s	PBX hostname or IP
--port, -p	SIP port (default: 5060)
--user, -u	SIP extension / username
--password	SIP password
--transport	udp, tcp, or tls (default: udp)
--srtp	disabled, optional, or mandatory (default: disabled)
--tls-verify	Verify TLS server certificate
--dest, -d	Destination phone number or SIP URI (required)
--timeout, -t	Call timeout in seconds (default: 60)

Audio source (mutually exclusive):

Flag	Description
--wav, -w	Path to WAV file to play (mutually exclusive with --interactive)
--interactive	Interactive live TTS mode: type text in the console that gets spoken via Piper TTS during the call (mutually exclusive with --wav, requires --piper-model)

TTS:

Flag	Description
--text	Text to synthesize via piper TTS, or initial greeting in interactive mode
--tts-model	Piper voice model for pre-generated TTS (default: de_DE-thorsten-high)
--piper-model	Path to Piper .onnx model for live TTS in interactive mode
--tts-sample-rate	Resample TTS output to this rate in Hz (default: native/22050)
--tts-data-dir	Directory for piper voice models (default: ~/.local/share/piper-voices)
--tts-cuda	Use CUDA GPU acceleration for Piper TTS

Playback timing:

Flag	Description
--pre-delay	Seconds to wait after answer before playback (default: 0)
--post-delay	Seconds to wait after playback before hangup (default: 0)
--inter-delay	Seconds to wait between WAV repeats (default: 0)
--repeat	Number of times to play the WAV (default: 1)
--wait-for-silence	Wait for N seconds of remote silence before playback (e.g. 1.0 to let callee finish "Hello?"). Uses SilenceDetector on the incoming audio RMS. Applied after --pre-delay.

Recording & transcription:

Flag	Description
--record	Record remote-party (RX) audio to this WAV file path (parent dirs created automatically)
--record-tx	Record local (TX) audio to this WAV file path
--mix-mode	Post-call mix mode: none, mono, or stereo (requires --record + --record-tx)
--mix-output	Output path for the RX+TX mix file
--transcribe	Transcribe recorded audio via STT and write a JSON call report (requires --record)
--live-transcribe	Live STT of remote audio during the call (real-time output to console)
--stt-backend	STT backend: faster-whisper (default) or openvino
--stt-model	Whisper model size for transcription (default: medium)
--stt-language	Language code for STT transcription (default: from config, then de)
--stt-data-dir	Directory for Whisper STT models (default: ~/.local/share/faster-whisper-models)

Live VAD (for --live-transcribe):

Flag	Description
--vad-silence-threshold	RMS silence threshold (default: 0.01)
--vad-silence-trigger	Seconds of silence to trigger chunk boundary (default: 0.3)
--vad-max-chunk	Max seconds per audio chunk (default: 5.0)
--vad-min-chunk	Min seconds per audio chunk (default: 0.5)

Audio device:

Flag	Description
--no-null-audio	Use real audio devices instead of null device for both directions
--real-capture	Use real microphone even when playback stays null
--real-playback	Use real speaker even when capture stays null
--play-audio	Play remote-party (RX) audio on local speakers via sounddevice
--play-tx	Route local (TX) audio to output sinks; stereo with RX when combined with --play-audio
--no-play-rx	Disable routing of RX audio to output sinks
--audio-device	Sounddevice output device (index or name substring)
--audio-socket	Unix socket path for live PCM streaming (16 kHz, S16_LE, mono)

NAT traversal:

Flag	Description
--stun-servers	Comma-separated STUN servers (e.g. stun.l.google.com:19302)
--ice	Enable ICE for media NAT traversal
--turn-server	TURN relay server (host:port)
--turn-username	TURN username
--turn-password	TURN password
--turn-transport	TURN transport: udp, tcp, or tls (default: udp)
--keepalive	UDP keepalive interval in seconds (0 = disabled)
--public-address	Public IP to advertise in SDP/Contact (e.g. K3s node IP)

Logging:

Flag	Description
--verbose, -v	Debug logging
--pjsip-log-level	PJSIP log verbosity (0–6, default: 3)

callee_autoanswer — Auto-Answer Incoming Calls

Answer incoming calls automatically with optional WAV playback or TTS:

# Auto-answer with no media (just answer and hold)
python -m sipstuff.cli callee_autoanswer \
    --server pbx.local --user 1001 --password secret

# Auto-answer and play a WAV file
python -m sipstuff.cli callee_autoanswer \
    --server pbx.local --user 1001 --password secret \
    --mode wav --wav-file greeting.wav

# Auto-answer with TTS greeting
python -m sipstuff.cli callee_autoanswer \
    --server pbx.local --user 1001 --password secret \
    --mode tts --tts-text "Hallo, bitte hinterlassen Sie eine Nachricht."

# Full sequence: start WAV → pause → TTS content → pause → end WAV
python -m sipstuff.cli callee_autoanswer \
    --server pbx.local --user 1001 --password secret \
    --mode tts --tts-text "Ihre Nachricht wird aufgezeichnet." \
    --start-wav ding.wav --end-wav bye.wav \
    --pause-before-content 0.5 --pause-before-end 1.0 \
    --answer-delay 2.0

Flag	Description
--mode	Playback mode: none, wav, or tts (default: none)
--wav-file	WAV file to play (requires --mode wav)
--tts-text	Text for TTS (requires --mode tts)
--piper-model	Piper voice model (default: de_DE-thorsten-high)
--tts-data-dir	Piper data directory
--start-wav	WAV file to play at call start
--end-wav	WAV file to play before hangup
--pause-before-start	Pause before start WAV (default: 0.0 s)
--pause-before-content	Pause before content WAV/TTS (default: 0.0 s)
--pause-before-end	Pause before end WAV (default: 0.0 s)
--answer-delay	Seconds before answering (default: 1.0)
--tts-cuda	Use CUDA GPU acceleration for Piper TTS
--no-auto-answer	Do not auto-answer calls

Shared flags: --config, --server, --port, --user, --password, --transport, --srtp, --tls-verify, --local-port, NAT traversal group, --verbose, --pjsip-log-level

callee_realtime-tts — Incoming Call with Live TTS

Answer incoming calls and speak text in real-time via Piper TTS:

# Answer and speak a greeting
python -m sipstuff.cli callee_realtime-tts \
    --server pbx.local --user 1001 --password secret \
    --tts-text "Willkommen, wie kann ich Ihnen helfen?" \
    --piper-model /path/to/de_DE-thorsten-high.onnx

# Interactive mode — type text during the call
python -m sipstuff.cli callee_realtime-tts \
    --server pbx.local --user 1001 --password secret \
    --interactive \
    --piper-model /path/to/de_DE-thorsten-high.onnx

TTS & playback:

Flag	Description
--tts-text	Initial TTS text spoken on answer
--interactive	Interactive mode: type text in the console that gets spoken live
--piper-model	Path to Piper .onnx model (default: ./de_DE-thorsten-high.onnx)
--tts-cuda	Use CUDA GPU acceleration for Piper TTS
--wav-file	WAV file to play at call start
--play-delay	Seconds to wait before playback (default: 0.0)
--answer-delay	Seconds before answering (default: 1.0)
--no-auto-answer	Do not auto-answer calls

STT (speech-to-text):

Flag	Description
--stt-backend	STT backend: faster-whisper (default) or openvino
--stt-model	Whisper model size or HuggingFace model ID (default: base)
--stt-live-model	Separate smaller model for live transcription
--stt-device	Compute device: cpu or cuda
--stt-language	Language code for STT
--stt-data-dir	Whisper model cache directory

Live VAD:

Flag	Description
--vad-silence-threshold	RMS silence threshold (default: 0.01)
--vad-silence-trigger	Seconds of silence to trigger chunk boundary (default: 0.3)
--vad-max-chunk	Max seconds per audio chunk (default: 5.0)
--vad-min-chunk	Min seconds per audio chunk (default: 0.5)

Recording & transcription:

Flag	Description
--wav-output	Save RX recording to this path
--wav-dir	Directory for WAV files (default: ..)
--wav-output-tx	Save TX recording to this path
--no-wav	Do not save WAV recordings
--transcribe	Full transcription of RX recording after call ends (writes JSON report)

Audio output:

Flag	Description
--audio-socket	Unix socket for live audio streaming (PCM 16 kHz, S16_LE, mono)
--play-audio	Play remote audio on local speakers via sounddevice
--audio-device	Sounddevice output device (index or name substring)

Shared flags: --config, --server, --port, --user, --password, --transport, --srtp, --tls-verify, --local-port, NAT traversal group, --verbose, --pjsip-log-level

callee_live-transcribe — Incoming Call with Live Transcription

Answer incoming calls and transcribe remote audio in real-time:

# Live transcription of incoming calls
python -m sipstuff.cli callee_live-transcribe \
    --server pbx.local --user 1001 --password secret \
    --stt-model base --stt-language de

# With recording and post-call full transcription
python -m sipstuff.cli callee_live-transcribe \
    --server pbx.local --user 1001 --password secret \
    --wav-output /tmp/rx.wav --transcribe \
    --stt-model medium

# With greeting WAV and audio streaming
python -m sipstuff.cli callee_live-transcribe \
    --server pbx.local --user 1001 --password secret \
    --wav-file greeting.wav --play-delay 0.5 \
    --audio-socket /tmp/audio.sock --play-audio

Flag	Description
--stt-backend	STT backend: faster-whisper (default) or openvino
--stt-model	Whisper model (default: base)
--stt-live-model	Separate smaller model for live transcription
--stt-device	Compute device: cpu or cuda
--stt-language	Language code for STT
--stt-data-dir	Whisper model cache directory
--vad-silence-threshold	RMS silence threshold (default: 0.01)
--vad-silence-trigger	Seconds of silence to trigger chunk boundary (default: 0.3)
--vad-max-chunk	Max seconds per audio chunk (default: 5.0)
--vad-min-chunk	Min seconds per audio chunk (default: 0.5)
--wav-output	Save RX recording to this path
--wav-output-tx	Save TX recording to this path
--wav-dir	Directory for WAV files (default: ..)
--no-wav	Do not save WAV recordings
--transcribe	Full transcription of RX recording after call ends (writes JSON report)
--wav-file	WAV file to play at call start
--tts-text	Text for Piper TTS playback at call start
--piper-model	Piper voice model (default: de_DE-thorsten-high)
--tts-data-dir	Piper data directory
--play-delay	Seconds to wait before playback (default: 0.0)
--audio-socket	Unix socket for live audio streaming (PCM 16 kHz, S16_LE, mono)
--play-audio	Play remote audio on local speakers via sounddevice
--audio-device	Sounddevice output device (index or name substring)
--answer-delay	Seconds before answering (default: 1.0)
--no-auto-answer	Do not auto-answer calls

Shared flags: --config, --server, --port, --user, --password, --transport, --srtp, --tls-verify, --local-port, NAT traversal group, --verbose, --pjsip-log-level

Docker / Podman Example

Convert a WAV file to 8 kHz mono PCM and place a TLS+SRTP call from a container:

ffmpeg -i alert.wav -ar 8000 -ac 1 -sample_fmt s16 -y /tmp/alert.wav 2>/dev/null && \
podman run --network=host -it --rm --userns=keep-id:uid=1200,gid=1201 \
    -v /tmp/alert.wav:/app/alert.wav:ro \
    xomoxcc/sipstuff:latest \
    python3 -m sipstuff.cli \
    call --server pbx.example.com \
    --port 5161 --transport tls --srtp mandatory \
    --user 1000 \
    --password changeme \
    --dest [phone] \
    --wav /app/alert.wav \
    --pre-delay 3.0 \
    --post-delay 1.0 \
    --repeat 3 -v

TTS call from a container (no WAV file needed on the host):

podman run --network=host -it --rm --userns=keep-id:uid=1200,gid=1201 \
    xomoxcc/sipstuff:latest \
    python3 -m sipstuff.cli \
    call --server pbx.example.com \
    --port 5161 --transport tls --srtp mandatory \
    --user 1000 \
    --password changeme \
    --dest [phone] \
    --text "Achtung! Wasserstand kritisch!" \
    --tts-sample-rate 8000 \
    --pre-delay 3.0 \
    --post-delay 1.0 \
    --repeat 3 -v

TTS with persistent voice models (avoids re-downloading on every --rm run):

podman run --network=host -it --rm --userns=keep-id:uid=1200,gid=1201 \
    -v ~/.local/share/piper-voices:/data/piper \
    xomoxcc/sipstuff:latest \
    python3 -m sipstuff.cli \
    call --server pbx.example.com \
    --port 5161 --transport tls --srtp mandatory \
    --user 1000 \
    --password changeme \
    --dest [phone] \
    --text "Achtung! Wasserstand kritisch!" \
    --tts-data-dir /data/piper \
    --tts-sample-rate 8000 \
    --pre-delay 3.0 -v

Passing connection parameters via environment variables:

podman run --network=host -it --rm --userns=keep-id:uid=1200,gid=1201 \
    -e SIP_SERVER=pbx.example.com \
    -e SIP_PORT=5161 \
    -e SIP_USER=1000 \
    -e SIP_PASSWORD=changeme \
    -e SIP_TRANSPORT=tls \
    -e SIP_SRTP=mandatory \
    xomoxcc/sipstuff:latest \
    python3 -m sipstuff.cli \
    call --dest [phone] \
    --text "Server nicht erreichbar!" \
    --tts-sample-rate 8000 -v

Callee auto-answer in a container:

podman run --network=host -it --rm --userns=keep-id:uid=1200,gid=1201 \
    -e SIP_SERVER=pbx.example.com \
    -e SIP_USER=1001 \
    -e SIP_PASSWORD=changeme \
    xomoxcc/sipstuff:latest \
    python3 -m sipstuff.cli \
    callee_autoanswer --mode tts \
    --tts-text "Hallo, bitte hinterlassen Sie eine Nachricht." -v

Notes:

• --network=host is needed for SIP/RTP media traffic.
• --userns=keep-id:uid=1200,gid=1201 maps the container's pythonuser to your host user (rootless Podman).
• The ffmpeg step in the WAV example ensures the file is in a SIP-friendly format (8 kHz, mono, 16-bit PCM).
• The TTS example needs no host-side WAV file — piper generates speech inside the container. Use --tts-sample-rate 8000 to resample for narrowband SIP.

Example Output

TTS call with silence detection, recording, and transcription:

TTS live transcribe with Cuda acceleration and live snddevice output from container:

Library Usage

from sipstuff import make_sip_call

# Simple call with a WAV file
success = make_sip_call(
    server="pbx.local",
    user="1000",
    password="secret",
    destination="[phone]",
    wav_file="/path/to/alert.wav",
    transport="udp",
    repeat=2,
)

# Call with TTS (no WAV file needed)
success = make_sip_call(
    server="pbx.local",
    user="1000",
    password="secret",
    destination="[phone]",
    text="Achtung! Wasserstand kritisch!",
    tts_model="de_DE-thorsten-high",
    pre_delay=1.0,
    post_delay=1.0,
)

# TLS + SRTP call with all options
success = make_sip_call(
    server="pbx.example.com",
    user="1000",
    password="secret",
    destination="[phone]",
    wav_file="/path/to/alert.wav",
    port=5061,
    transport="tls",
    timeout=30,
    pre_delay=1.5,
    post_delay=2.0,
    inter_delay=1.0,
    repeat=3,
)

# Play remote audio on local speakers
success = make_sip_call(
    server="pbx.local",
    user="1000",
    password="secret",
    destination="[phone]",
    wav_file="alert.wav",
    play_audio=True,
)

Context Manager for Multiple Calls

from sipstuff import SipCaller, SipCallerConfig, WavPlayConfig

config = SipCallerConfig.from_config(config_path="sip_config.yaml")
with SipCaller(config) as caller:
    caller.make_call("[phone]", wav_play=WavPlayConfig(wav_path="alert.wav"))
    caller.make_call("[phone]", wav_play=WavPlayConfig(wav_path="other.wav"))

Using TTS Directly

from sipstuff import generate_wav, TtsError

# Generate a WAV file from text (auto-downloads model on first use)
try:
    wav_path = generate_wav(
        text="Server nicht erreichbar!",
        model="de_DE-thorsten-high",
        sample_rate=8000,  # resample for narrowband SIP (0 = keep native)
    )
    # wav_path is a temporary file — use it, then clean up
    print(f"Generated: {wav_path}")
except TtsError as exc:
    print(f"TTS failed: {exc}")

# With a specific output path and model directory
wav_path = generate_wav(
    text="Wasserstand kritisch!",
    output_path="/tmp/alert_tts.wav",
    data_dir="/opt/piper-voices",
)

Using STT (Speech-to-Text) Directly

from sipstuff import transcribe_wav, SttError

# Transcribe a recorded call (auto-downloads model on first use)
# VAD (Silero Voice Activity Detection) is enabled by default —
# recommended for phone recordings with silences/ringing tones.
try:
    text, meta = transcribe_wav("/tmp/recording.wav")  # default: medium model, German, vad=True
    print(f"Transcription: {text}")
    print(f"Duration: {meta['audio_duration']:.1f}s, segments: {len(meta['segments'])}")
    for seg in meta["segments"]:
        print(f"  [{seg['start']:.1f}s - {seg['end']:.1f}s] {seg['text']}")
except SttError as exc:
    print(f"STT failed: {exc}")

# English transcription with a different model, VAD disabled
text, meta = transcribe_wav("/tmp/recording.wav", language="en", model="large-v3", vad_filter=False)

# Custom model cache directory (useful for Docker volumes)
text, meta = transcribe_wav(
    "/tmp/recording.wav",
    data_dir="/opt/whisper-models",
)

Call with Recording, Silence Detection, and Transcription

from sipstuff import SipCaller, SipCallerConfig, WavPlayConfig, RecordingConfig, transcribe_wav

config = SipCallerConfig.from_config(config_path="sip_config.yaml")
with SipCaller(config) as caller:
    success = caller.make_call(
        "[phone]",
        wav_play=WavPlayConfig(wav_path="alert.wav"),
        recording=RecordingConfig(record_path="/tmp/recording.wav"),
    )
    if success:
        text, meta = transcribe_wav("/tmp/recording.wav")
        print(f"Remote party said: {text}")
        print(f"Segments: {meta['segments']}")

Error Handling

from sipstuff import make_sip_call, SipCallError, TtsError, SttError

try:
    success = make_sip_call(
        server="pbx.local",
        user="1000",
        password="secret",
        destination="[phone]",
        wav_file="alert.wav",
    )
    if not success:
        print("Call was not answered")
except SipCallError as exc:
    print(f"SIP error: {exc}")  # registration, transport, or WAV issues
except TtsError as exc:
    print(f"TTS error: {exc}")  # piper not found, synthesis failed
except SttError as exc:
    print(f"STT error: {exc}")  # faster-whisper not found, transcription failed

Configuration

Configuration is loaded with the following priority (highest first): CLI flags / overrides > environment variables > YAML file.

YAML Config

See example_config.yaml for a full example:

sip:
  server: "pbx.example.com"
  port: 5060
  user: "1000"
  password: "changeme"
  transport: "udp"        # udp, tcp, or tls
  srtp: "disabled"        # disabled, optional, or mandatory
  tls_verify_server: false
  local_port: 0           # 0 = auto

call:
  timeout: 60
  pre_delay: 0.0
  post_delay: 0.0
  inter_delay: 0.0
  repeat: 1
  wait_for_silence: 0.0    # seconds of remote silence before playback (0 = disabled)

tts:
  model: "de_DE-thorsten-high"  # piper voice model (auto-downloaded on first use)
  sample_rate: 0                # 0 = keep native (22050), use 8000 for narrowband SIP

NAT Traversal

For hosts behind NAT, add a nat: section to the YAML config:

nat:
  stun_servers:                # STUN servers for public IP discovery
    - "stun.l.google.com:19302"
  stun_ignore_failure: true    # continue if STUN unreachable (default: true)
  ice_enabled: true            # ICE for media NAT traversal
  turn_enabled: false          # TURN relay (requires turn_server)
  turn_server: ""              # host:port
  turn_username: ""
  turn_password: ""
  turn_transport: "udp"        # udp, tcp, or tls
  keepalive_sec: 15            # UDP keepalive interval (0 = disabled)
  public_address: ""           # Manual public IP for SDP/Contact (e.g. K3s node IP)

• STUN only — sufficient when the NAT gateway preserves port mappings (cone NAT). Lets PJSIP discover its public IP for the SDP c= line.
• STUN + ICE — adds connectivity checks so both sides probe multiple candidate pairs. Handles most residential/corporate NATs.
• TURN — needed when the NAT is symmetric or a firewall blocks direct UDP. All media is relayed through the TURN server, adding latency but guaranteeing connectivity.
• Keepalive — sends periodic UDP packets to keep NAT bindings alive during long calls or idle registrations.

Static NAT / K3s

When running in a K3s pod (10.x.x.x pod IP) calling a SIP server in a DMZ (e.g. Fritzbox), the auto-detected local IP is the pod IP which is unreachable from the SIP server. STUN doesn't help either — it returns the WAN IP, not the node IP. Use public_address to manually set the IP that appears in SDP and SIP Contact headers, while the socket stays bound to the pod IP:

nat:
  public_address: "192.168.1.50"  # K3s node IP reachable from the SIP server

Or via CLI / environment variable:

# CLI
python -m sipstuff.cli call --public-address 192.168.1.50 --dest [phone] --wav alert.wav

# Environment variable
export SIP_PUBLIC_ADDRESS=192.168.1.50

This works because K3s uses SNAT (conntrack) for pod-to-external traffic — reply packets from the Fritzbox to the node IP are translated back to the pod IP by the kernel's connection tracking.

Environment Variables

All settings can be set via SIP_ prefixed environment variables:

SIP connection:

Variable	Maps to
SIP_SERVER	sip.server
SIP_PORT	sip.port
SIP_USER	sip.user
SIP_PASSWORD	sip.password
SIP_TRANSPORT	sip.transport
SIP_SRTP	sip.srtp
SIP_TLS_VERIFY_SERVER	sip.tls_verify_server
SIP_LOCAL_PORT	sip.local_port

Call timing:

Variable	Maps to
SIP_TIMEOUT	call.timeout
SIP_PRE_DELAY	call.pre_delay
SIP_POST_DELAY	call.post_delay
SIP_INTER_DELAY	call.inter_delay
SIP_REPEAT	call.repeat
SIP_WAIT_FOR_SILENCE	call.wait_for_silence

TTS:

Variable	Maps to
SIP_TTS_MODEL	tts.model
SIP_TTS_SAMPLE_RATE	tts.sample_rate
SIP_TTS_CUDA	tts.use_cuda
SIP_TTS_DATA_DIR	tts.data_dir

NAT traversal:

Variable	Maps to
SIP_STUN_SERVERS	nat.stun_servers (comma-separated)
SIP_STUN_IGNORE_FAILURE	nat.stun_ignore_failure
SIP_ICE_ENABLED	nat.ice_enabled
SIP_TURN_ENABLED	nat.turn_enabled
SIP_TURN_SERVER	nat.turn_server
SIP_TURN_USERNAME	nat.turn_username
SIP_TURN_PASSWORD	nat.turn_password
SIP_TURN_TRANSPORT	nat.turn_transport
SIP_KEEPALIVE_SEC	nat.keepalive_sec
SIP_PUBLIC_ADDRESS	nat.public_address

Audio device:

Variable	Maps to
SIP_NULL_AUDIO	audio.use_null_audio (default: true)
SIP_NULL_CAPTURE	audio.null_capture (null mic; None inherits SIP_NULL_AUDIO)
SIP_NULL_PLAYBACK	audio.null_playback (null speaker; None inherits SIP_NULL_AUDIO)
SIP_PLAY_RX	audio.play_rx (route RX to output sinks; default: true)
SIP_PLAY_TX	audio.play_tx (route TX to output sinks; default: false)
SIP_AUDIO_DEVICE	audio.audio_device (sounddevice index or name)

STT:

Variable	Maps to
SIP_STT_BACKEND	stt.backend
SIP_STT_MODEL	stt.model
SIP_STT_LANGUAGE	stt.language
SIP_STT_DEVICE	stt.device
SIP_STT_DATA_DIR	stt.data_dir
SIP_LIVE_TRANSCRIBE	stt.live_transcribe

VAD:

Variable	Maps to
SIP_VAD_SILENCE_THRESHOLD	vad.silence_threshold
SIP_VAD_SILENCE_TRIGGER	vad.silence_trigger
SIP_VAD_MAX_CHUNK	vad.max_chunk
SIP_VAD_MIN_CHUNK	vad.min_chunk

Callee:

Variable	Maps to
SIP_AUTO_ANSWER	callee.auto_answer (default: true)
SIP_ANSWER_DELAY	callee.answer_delay (default: 1.0)

TTS runtime:

Variable	Default	Description
PIPER_DATA_DIR	~/.local/share/piper-voices	Directory for downloaded voice models

STT runtime (faster-whisper):

Variable	Default	Description
WHISPER_DATA_DIR	~/.local/share/faster-whisper-models	Directory for downloaded Whisper models
WHISPER_MODEL	medium	Default model size (tiny, base, small, medium, large-v3)
WHISPER_DEVICE	cpu	Compute device (cpu or cuda)
WHISPER_COMPUTE_TYPE	int8 (CPU) / float16 (CUDA)	Quantization type (int8, float16, float32)

PJSIP native log routing:

Variable	Default	Description
PJSIP_LOG_LEVEL	3	PJSIP log verbosity routed through loguru (0 = none … 6 = trace)
PJSIP_CONSOLE_LEVEL	4	PJSIP native console output printed directly to stdout (4 = PJSIP default, 0 = suppressed)

All PJSIP native output is captured by a pj.LogWriter subclass and forwarded to loguru (classname="pjsip"). PJSIP_LOG_LEVEL controls what the writer receives; PJSIP_CONSOLE_LEVEL controls what PJSIP additionally prints to stdout on its own. Set PJSIP_CONSOLE_LEVEL=0 to suppress native output and rely solely on loguru.

Both values can also be passed directly to the SipCaller constructor (pjsip_log_level=, pjsip_console_level=), which takes highest priority over env vars and class defaults.

Troubleshooting

One-way audio on multi-homed hosts

On hosts with multiple network interfaces PJSIP may auto-detect the wrong local IP and advertise it in the SDP c= line. The SIP gateway then sends RTP to the wrong address, resulting in one-way audio (you can see RX total 0pkt in the call stats). sip_endpoint.py works around this by calling _local_address_for() at startup — a non-sending UDP connect to the SIP server that lets the kernel's routing table select the correct source address. Both the SIP transport and the account media transport are then bound to that IP.

Symptom in logs:

pjsua_acc.c !....IP address change detected for account 0 (192.168.x.x --> 192.168.y.y)

PJSIP_ETPNOTSUITABLE warning on INVITE

Temporary failure in sending Request msg INVITE ... Unsuitable transport selected (PJSIP_ETPNOTSUITABLE)

This is a PJSIP-internal transport selection warning that appears when the SIP URI includes ;transport=udp and the bound transport doesn't match PJSIP's initial candidate. The call still succeeds on the retry. It does not affect media or audio quality.

conference.c Remove port failed warning on hangup

conference.c !.Remove port failed: Invalid value or argument (PJ_EINVAL)

A cosmetic PJSIP warning during WAV player cleanup. The player's conference port is invalidated when the call's media is torn down; the subsequent C++ destructor tries to remove it again. Audio and call lifecycle are unaffected.

WAV File Requirements

The module accepts standard WAV files. Recommended format for SIP:

• 16-bit PCM, mono, 8000 or 16000 Hz sample rate

Non-standard formats (stereo, different bit depths/rates) will produce warnings but playback is still attempted.

Module Structure

File	Purpose
__init__.py	Public API: make_sip_call, SipCaller, SipCallee, SipEndpoint, SipCallError, CallResult, config classes, generate_wav, transcribe_wav, configure_logging, print_banner
sip_endpoint.py	SipEndpoint base class for PJSUA2 Endpoint lifecycle, SipCaller(SipEndpoint) with make_call() orchestration, _PjLogWriter, _local_address_for()
sip_call.py	SipCall(pj.Call) shared base, SipCallerCall (caller-side media management), SipCalleeCall (callee-side hooks)
sip_media.py	SilenceDetector, AudioStreamPort, TranscriptionPort — PJSUA2 AudioMediaPort subclasses
sip_account.py	SipAccount(pj.Account) base, SipCallerAccount (incoming-call rejection), SipCalleeAccount (auto-answer dispatch)
sip_callee.py	SipCallee(SipEndpoint) — incoming call handling orchestration
sip_types.py	CallResult dataclass, SipCallError, WavInfo
sip_caller.py	Legacy module (retained for imports)
sipconfig.py	Pydantic v2 config models (SipEndpointConfig, SipCallerConfig, SipCalleeConfig) with YAML / env / override loading
vad.py	VADAudioBuffer — voice activity detection buffer for live transcription
audio.py	resample_linear() (numpy), ensure_wav_16k_mono() — audio format utilities
snddevice_list.py	Sound device enumeration utility
tts/tts.py	Piper TTS integration: text-to-WAV generation via Python API with optional resampling
tts/live.py	Live TTS streaming: PiperTTSProducer (producer thread), TTSMediaPort (PJSIP consumer), interactive_console()
stt/stt.py	Speech-to-text via faster-whisper: WAV transcription with Silero VAD pre-filtering and segment timestamps
stt/live.py	Live STT streaming during calls
autoanswer/	Callee auto-answer implementation (integrated as callee_autoanswer CLI subcommand)
realtime/	Callee real-time TTS implementation (integrated as callee_realtime-tts CLI subcommand)
transcribe/	Callee live-transcribe implementation (integrated as callee_live-transcribe CLI subcommand)
training/	Voice training utilities: record_dataset.py, record_gui.py (standalone, not in CLI) — see training README

| cli.py | CLI entry point with six subcommands | | example_config.yaml | Sample configuration file | | dist_scripts/install_pjsip.sh | Build script for PJSIP with Python bindings (default: 2.16) |

Dockerfile (Two-Stage Build)

1. pjsip-builder (python:3.14-slim-trixie): builds PJSIP from source, copies .so libs + Python SWIG bindings
2. Main image (python:3.14-slim-trixie): copies PJSIP artifacts, installs sipstuff (including piper-tts via pip). Runs as non-root pythonuser (UID 1200). Locale de_DE.UTF-8, tz Europe/Berlin. Entrypoint: tini.

Optional build args:

• INSTALL_CUDA=true: installs NVIDIA CUDA runtime libs for faster-whisper GPU inference
• INSTALL_OPENVINO=true: installs optimum-intel[openvino] for OpenVINO STT backend (Intel GPU/CPU)

# Standard build
docker build -t xomoxcc/sipstuff:latest .

# With CUDA support
docker build --build-arg INSTALL_CUDA=true -t xomoxcc/sipstuff:latest .

# With OpenVINO support
docker build --build-arg INSTALL_OPENVINO=true -t xomoxcc/sipstuff:latest .

Development

make install          # Create .venv (Python 3.14), install deps
make tests            # pytest .
make lint             # black .
make isort            # isort .
make tcheck           # mypy . (strict mode + pydantic plugin)
make commit-checks    # pre-commit run --all-files (black --check, mypy, gitleaks)
make prepare          # tests + commit-checks
make build            # docker build
make pypibuild        # hatch build --clean

Tests

pytest tests/ -v

Tests cover config validation (SipConfig, CallConfig, SipCallerConfig, NatConfig), config loading (YAML, env vars, overrides priority), WAV validation, and mocked PJSUA2 caller behavior including endpoint lifecycle, context manager, STUN/ICE/TURN config application, and public_address handling. No real SIP server needed.

CI Pipeline (GitHub Actions)

checkblack.yml → mypynpytests.yml → buildmultiarchandpush.yml (amd64+arm64, pushes to DockerHub)

CI skips venv setup when GITHUB_RUN_ID is set — installs deps directly.

License

This project is licensed under the LGPL where applicable/possible — see LICENSE.md. Some files/parts may use other licenses: MIT | GPL | LGPL. Always check per‑file headers/comments.

Authors

• Repo owner (primary author)
• Additional attributions are noted inline in code comments

Acknowledgments

• Inspirations and snippets are referenced in code comments where appropriate.

⚠️ Note

This is a development/experimental project. For production use, review security settings, customize configurations, and test thoroughly in your environment. Provided "as is" without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the software or the use or other dealings in the software. Use at your own risk.