ft8lib 1.2.0

FT8 and FT4 amateur-radio digital mode encoder/decoder (Python port of WSJT-X)

ft8lib

FT8 and FT4 amateur-radio digital modes in Python.

ft8lib is a Python port of the FT8/FT4 protocol implementation from WSJT-X: message packing (77-bit), CRC-14, LDPC(174,91) forward error correction, GFSK waveform synthesis, and a full receive chain (candidate sync search, downconversion, fine time/frequency sync, noncoherent multi-symbol soft demodulation, belief-propagation decoding).

Installation

pip install .

The only runtime dependency is numpy. The decode hot paths (LDPC belief propagation, ordered-statistics decoding, fine sync) are compiled from a small C extension at install time; if the extension is unavailable the library transparently falls back to slower pure-numpy implementations of the same algorithms.

Usage

Python API

import numpy as np
import ft8lib

# --- Encode a message to an audio waveform (float64, 12000 S/s) ---
wave = ft8lib.encode_ft8("CQ K1ABC FN42", f0=1500.0)   # 12.64 s
wave4 = ft8lib.encode_ft4("W9XYZ K1ABC RR73", f0=1500.0)  # 5.04 s

# Transmission conventionally starts 0.5 s into the 15 s (FT8) or
# 7.5 s (FT4) cycle:
period = np.zeros(ft8lib.FT8.NMAX)
period[6000:6000 + len(wave)] = wave

# --- Decode a receive period (12000 S/s audio, any amplitude scale) ---
for result in ft8lib.decode_ft8(period):
    print(result)
# FT8 -15 dB  DT +0.00 s  1500.0 Hz  CQ K1ABC FN42

for result in ft8lib.decode_ft4(audio_7p5s):
    print(result)

Each decode is a ft8lib.Decode with fields message, snr (dB in 2500 Hz), dt (seconds relative to the nominal 0.5 s start), freq (Hz), sync, and mode.

Realtime decoding

To decode live audio, feed a RealtimeDecoder blocks of samples as they arrive (starting at a cycle boundary). Like WSJT-X, it runs decode attempts part-way through the receive period (FT8: 11.8, 13.5 and 14.7 s; FT4: 5.6 and 6.05 s), so most messages are reported seconds before the period ends, and it rolls over automatically at each cycle boundary:

rt = ft8lib.RealtimeDecoder("FT8", mycall="K1ABC")
for block in audio_blocks:            # e.g. sound-card capture callback
    for result in rt.feed(block):     # new decodes, as soon as they appear
        print(rt.cycle, result)
rt.flush()                            # decode a partial final period

ft8lib.decode_realtime(blocks, mode="FT8") wraps this as a generator over an iterable of blocks. For a single already-captured period there are also decode_ft8_stream / decode_ft4_stream, generator versions of the batch decoders that yield each message as soon as it decodes instead of returning the full list at the end.

Lower-level building blocks are exposed too:

bits = ft8lib.pack77("CQ K1ABC FN42")        # 77-bit numpy array
msg, ok = ft8lib.unpack77(bits)              # back to text
tones = ft8lib.ft8_tones_from_bits(bits)     # 79 channel symbols (0-7)
codeword = ft8lib.encode174_91(bits)         # 174-bit LDPC codeword

Hashed nonstandard callsigns (<PJ4/K1ABC>) resolve through a ft8lib.HashTable, which you can keep across decode cycles and prime with table.set_station(mycall, dxcall) — the same behavior as WSJT-X.

Command line

# Encode to a WAV file (12 kHz, 16-bit mono)
ft8lib encode "CQ K1ABC FN42" cq.wav --mode ft8 --freq 1500 --full-period

# Decode a recording (any sample rate; resampled to 12 kHz)
ft8lib decode 210701_133000.wav --mode ft8

Message types supported

Standard messages (types 1/2 incl. /R and /P), free text (0.0), DXpedition mode (0.1), ARRL Field Day (0.3/0.4), telemetry (0.5), WSPR-style (0.6), ARRL RTTY roundup (3), nonstandard/hashed calls (4), and EU VHF contest (5).

Performance

Decode thresholds measured with additive white Gaussian noise (10 trials per point, single signal):

Mode	~100% decode down to	WSJT-X (with OSD+AP)
FT8	−19 dB	−21 dB
FT4	−16 dB	−17.5 dB

The port implements the belief-propagation decoder with all five WSJT-X metric passes, but not ordered-statistics decoding (OSD), a-priori (AP) decoding, or multi-pass signal subtraction, which account for the remaining 1–2 dB. A full 15 s FT8 period with a busy band decodes in well under a second on a typical desktop.

Development

Detailed docs on the encode and decode pipelines, with Fortran-source correspondence tables, live in docs/encode.md and docs/decode.md.

The LDPC tables in src/ft8lib/_tables.py are generated from the WSJT-X Fortran sources by tools/gen_tables.py (expects the wsjtx source tree in the repository root).

Run the tests with:

pip install -e .[test]
pytest

License

GPL-3.0-or-later. This library is a derivative work of WSJT-X, Copyright (C) 2001-2024 by Joe Taylor, K1JT, and the WSJT-X Development Team, licensed under the GNU GPL v3. See LICENSE.