TCP/IP over a soundcard, up to ~2 Mbit/s aggregate full-duplex

A software modem that turns two PCs' soundcards into a point-to-point network link. It modulates IP packets onto audio with OFDM (DMT), bonds the left and right stereo channels into two parallel lanes, protects everything with Reed-Solomon FEC and an ARQ with fast retransmit, and hands packets to the OS through various virtual interfaces. Written in pure Python + NumPy.

Click images to expand

pip install numpy sounddevice         # sounddevice needs PortAudio
pip install pyserial                  # only required for --hostif serial
sudo apt install python3-tk           # only required for the UI on Linux

Run the bundled regression suite:

Test 14 simulated-cable scenarios in about 90 seconds without a soundcard:

python -m audiomodem --selftest

Connect two stereo 3.5 mm TRS cables, crossed.

Each PC's line-out must go into the other PC's line-in.

Bonding uses both left and right channels, so do not use mono cables.

For the 192k profile, set both machines' playback and capture devices to:

24-bit, 192000 Hz

Windows:

Sound Control Panel → Device → Advanced → Default Format

Linux:

ALSA usually follows requested rates automatically. PipeWire/Pulse users may need to allow 192 kHz in the daemon config.

CRITICAL: Disable all audio enhancements on both ends:

• Loudness EQ
• Noise suppression
• Echo cancellation
• AGC
• Spatial sound

These will destroy the waveform.

Set levels to about 80%.

The RX level meter should sit around:

-20 to -10 dBFS

On PC 1:

python -m audiomodem --fs 192000

The log should show something like:

host interface: pty /dev/pts/4 (slip)

Then run:

sudo slattach -L -p slip /dev/pts/4
sudo ifconfig sl0 10.0.0.1 pointopoint 10.0.0.2 mtu 1500 up

On PC 2:

Repeat the same commands with the IPs swapped:

sudo slattach -L -p slip /dev/pts/4
sudo ifconfig sl0 10.0.0.2 pointopoint 10.0.0.1 mtu 1500 up

Wait for LINK, then you are ready to ping, SSH, or SCP.

--hostif tun      Real IP interface, needs root
--hostif tcp      KISS over TCP for TNC software
--hostif serial   Windows + com0com virtual COM pair
--no-ui           Headless operation

Bundled simulation, both directions at once:

A real direct cable through decent onboard codecs typically measures 30–45 dB.

Expect 256-QAM or 1024-QAM as the standard operating point.

--mode auto is the default. It climbs as measured SNR allows and backs off on retransmission pressure.

Modulation: DMT/OFDM, 512-point FFT, 64-sample cyclic prefix.

Carriers: 189 carriers to 18 kHz at 48k. 229 carriers from 3 kHz to 87.4 kHz at 192k.

Synchronization: 24 pilots track phase and timing, driving a closed-loop sample-rate-offset corrector.

The modem uses streaming 8x polyphase plus windowed-sinc fractional resampling ahead of the demodulator.

This nulls the sample-rate offset, allowing even 4096-QAM to survive drifting clocks up to ±300 ppm.

Framing and Mapping: Schmidl-Cox sync, LS channel-estimation preamble, rep-3 QPSK header, and up to 24 data symbols at 48k or 32 data symbols at 192k.

Gray-mapped QPSK through 4096-QAM per symbol.

Data Integrity: RS(255,223), interleaver, scrambler, and CRC-32.

Link Layer: Go-back-N with window 8 or 16, piggybacked cumulative ACKs, duplicate-ACK fast retransmit, serialization-aware adaptive RTO, and a receive reorder buffer that absorbs bonded-lane skew.

Check cable directions and volumes.

The RX level must actively move.

Check the reverse cable.

SNR is too low for the current mode, or an OS audio enhancement is silently mangling the audio.

The codec or OS device format is not correctly set to 192 kHz on both ends.

Fall back to:

--fs 48000

Check both SYNC LEDs to make sure bonding is active.

Also verify that the constellation map looks like a distinct grid, not a blurry cloud.