James Hansen

FSVR, the Formant Synthesizer Virtual Rack, is a software reconstruction of the Yamaha FS1R, a 1998 rack synthesizer. It runs as a VST3, a CLAP and a standalone on Windows, macOS and Linux, with all 88 FM algorithms, the 1408 factory voices, 384 performances and 90 formant sequences read out of the ROM, and the FS1R's own front panel as the interface. Free and open source under the GPLv3.

This is a collaborative project. In 2025 Zhiyuan Wan (rgwan) started taking a real FS1R apart, reading the SH7044's 256 KB internal flash out through a hooked UART, dumping the 2 MB EPROM and the PLG150-DX ROM, capturing the board's own UART traffic, and publishing all of it as rgwan/fs1r_firmware_RE with a Ghidra project that already had the boot, loader, flash and LCD code named. Progress went up along the way on the yamahamusicians.com forum. None of this exists without that work. I found it in September 2026, started reading the decompiled firmware to see whether the CPU side could be rewritten rather than emulated, and opened this repository. We work on it together now: Zhiyuan runs a patched firmware on the unit to poke memory and read registers back, and records the hardware measurements the engine is calibrated against. The repository passed 20 stars in its first week, and Zhiyuan and I coordinate through the musica.studio Discord.

I have composed music with FM synthesis for a long time, my initial interest starting with the music on the SEGA Genesis/Mega Drive console, which used FM synthesis on its primary sound chip. My interest in this particular synth was piqued in 2010, when I first saw musictrackjp's demo and review of it, below. The FS1R is the only Yamaha synth that pairs eight-operator FM with formant operators, which shape a spectral window instead of putting out a sine, and that is why nothing else sounds like it. It was built for two years, its tone generator (the YMP706) shipped in exactly two products and was never documented, and a working unit now costs more than most people will spend on a box with a two-line display.

FSVR is a rewrite, not an emulation of the board. The SH7044 is not emulated and its firmware does not run here. The note-on path, the 192.3 Hz tick pipeline, the parameter conversion tables and the 88-algorithm routing table are read straight out of the decompiled firmware and ported to C++, so patches, performances and formant sequences load and play with the same parameter interpretation the hardware uses, because the same logic computes them. Nobody has register semantics for the two custom chips, so those stay models, and every constant behind them sits in one namespace in the source where calibrating against a recording of a real unit is a single table edit.

Much like FM8.plus, the tooling is why this is possible at all. The firmware was disassembled with Ghidra into a database a coding agent can query, and the first working recreation was written with Claude Fable 5.1 reading 1998 SH-2 firmware at something close to source level. The analysis and engine work have carried on with each model since, and anyone working on the project can reach the same tools through the Discord. rgwan ran a debug monitor over the whole capture set on the unit and read the voice image back segment by segment, 240,476 checks across 802 segments with no mismatch, so the CPU side is verified byte for byte against the hardware.

Most of that work happens in one shared agent environment rather than on either of our machines. Hermes Agent runs Claude Fable 5.1 in a container that holds both repositories, the Ghidra database and the capture tooling, and it takes direction from a channel on the Discord, so Zhiyuan and I are giving instructions to the same agent from different countries. Zhiyuan has the hardware, plays a capture request into the real FS1R and reports where FSVR diverges from it; I manage the project, design the frontend and developed tooling for managing the disassembly; the agent does the reading, the edits and the verification in between. It also keeps its own notes: when a technique costs it several attempts, it writes the recipe back into its skill files so the next session starts from it. This is one exchange, condensed.

Built in C++17 with CMake and JUCE, with Python tooling for the Ghidra work, the table and preset extraction, and the hardware capture and analysis scripts. Every push builds the VST3, the CLAP and the standalone for all three platforms through GitHub Actions, so a build for an OS you do not own is one Actions run away. No Yamaha firmware or ROM image ships in the repository, and the tools that regenerate every extracted table from a dump you make yourself are in it. Yamaha and FS1R are trademarks of Yamaha Corporation, and the project is not affiliated with them. What is left is fidelity work: the effects, the unvoiced pedestal above 4 kHz, AM sensitivity at deep modulation, and the last decibel of the filter envelope.