Design

TinyModem is header only: src/TinyModem.h includes everything, the modems are in src/modem/. All audio and DSP functionality is based on the arduino-audio-tools (codecs, streams, sine generators, Goertzel filters, Reed-Solomon, tasks and mutexes).

Class structure

Each modem consists of a configuration, an encoder and a decoder:

• The encoders are derived from ModemEncoder, the decoders from ModemDecoder (both in ModemBase.h), which are AudioTools AudioEncoder and AudioDecoder. The base classes define the common API (endTransmission(), writeIdle(), isCarrier(), isFinished(), the status output, …).
• Shared helpers: ToneOutput writes the tones of a SoundGenerator (FSK, MFSK, AX.25), SampleOutput writes calculated samples (PSK, OFDM, ggwave), PcmInput assembles the PCM frames and extracts the selected channel, and ByteQueue is the receive buffer of the streams.
• The streams are templates: BasicModemStream<Encoder, Decoder, Config> provides the Arduino Stream API, and BasicModemStreamRTOS<Base> adds the receive task (AudioTools Task and Mutex). ModemStream, MFSKStream, PSKStream, OFDMStream, AX25Stream, GGWaveStream and their RTOS versions are the concrete classes.
• PacketStream works on any Stream, independent of the modems.
• The command line tool works with ModemEncoder and ModemDecoder, so the same code supports all modems.

All decoders work with push processing: write() accepts PCM data in any chunk size, and the decoding is done whenever enough samples are buffered. flush() / end() pad the input with silence so that the last transmission is completed. The streams call write() from available() / read(), or from a separate task in the RTOS versions. The receive queue grows when needed, so no decoded data is lost: the input is only read while less than rx_buffer_size bytes are waiting.

The algorithms of the extensions are described in their documents: MFSK, PSK, OFDM, AX.25, ggwave and packets.

FSK: rewrite of minimodem

The FSK modem is a C++ rewrite of minimodem 0.24. The modem algorithms are unchanged, and the platform layer is replaced by the arduino-audio-tools.

minimodem	TinyModem
minimodem.c option defaults and baud modes	ModemConfig::setMode(), ModemConfig::resolve()
minimodem.c transmit (fsk_transmit_frame, leader, sync, trailer)	FSKEncoder
minimodem.c receive main loop	FSKDecoder::step()
fsk.c (frame search, confidence, carrier detection)	FSKDemodulator
simple-tone-generator.c	FSKToneGenerator (AudioTools FloatSineGenerator), replaceable by any phase continuous SoundGenerator
databits_*.c, baudot.c, uic_codes.c	DataCodecs.h
simpleaudio* (ALSA, PulseAudio, sndio, sndfile)	AudioTools streams. The CLI uses WAVEncoder, WAVDecoder and PortAudioStream.
FFTW	AudioTools GoertzelDetector

Goertzel instead of FFT

minimodem computes a full FFT for every bit, but only reads two bins: mark and space. This rewrite computes exactly those two bins with Goertzel filters. The frequency of band b is b * sample_rate / fftsize, the same FFT bin center the original uses, and the magnitude is scaled by 2 / nsamples. The results match the original (see the confidence values in the interop tests), and FFTW isn't needed.

Carrier auto-detection scans every band with a Goertzel filter. That is O(bands × samples) instead of one FFT, so it is slower than the original.

The GoertzelDetector of older arduino-audio-tools versions computed double precision sin() and cos() at the end of each block, which is once per analysed bit here. That is slow on MCUs without a double precision FPU (ESP32, RP2040), so recent versions compute them once in begin().

From a blocking loop to push processing

The original receive loop blocks in simpleaudio_read() until half the sample buffer is filled. FSKDecoder::write() runs one iteration of the original loop (step()) whenever enough samples for a complete frame search are buffered.

The transmitter's interactive behaviour (end of transmission after one idle bit time, idle carrier) used SIGALRM and select(). In the library it is an explicit endTransmission() / writeIdle(). The CLI emulates the original timing with select().

All state that was global in the C code is now kept in class instances, so you can use several encoders and decoders at the same time. This includes the Baudot shift state, the Caller-ID buffer and the tone generator phase.

Known differences

• The decoder doesn't reach exactly confidence=inf for perfect signals (float precision of the Goertzel filter). Decoding results are not affected.
• The generated audio differs from the original by at most 31 (of 32767) per sample: the original uses a 4096-entry sine table, this rewrite uses sinf().
• The AudioTools WAVDecoder drops the last incomplete batch of converted 24-bit or float samples in end(). The CLI works around this by appending silence.
• RTTY/TDD transmit: each transmission starts with a LTRS or FIGS code, because the receiver resets its shift state with each new carrier. The original only sends a shift code when the character set changes, so a second transmission that continues with figures was decoded as letters. The first transmission is identical to the original.
• Bugs fixed: a possible overflow of the 64-entry bit arrays, the unchecked SDMF Caller-ID message length, and a DC offset instead of noise in --Xrxnoise.