dx7toFM1

Split a Yamaha DX7 32-voice bank (.syx) into individual single-voice .syx files, each named after the voice it contains, so FM1's third-party firmware can import them one by one.

Why this exists

FM1 plays DX7 voices, but its third-party firmware only imports one voice per file — it has no support for banks. What actually circulates online — factory ROMs, cartridge rips, patch collections — is almost always a 32-voice bank dump, which the firmware cannot use as-is:

	DX7 32-voice bank	DX7 single voice
Header	`F0 43 0n 09 20 00`	`F0 43 0n 00 01 1B`
Payload	4096 bytes — 128 per voice, packed	155 bytes, unpacked
Total file size	4104 bytes	163 bytes
FM1 third-party firmware	❌ cannot import	✅ imports directly

The two are not just different sizes: in a bank, each voice is bit-packed into 128 bytes, with several parameters sharing the same byte. A single-voice dump stores one parameter per byte in 155 bytes. So a bank has to be unpacked and re-wrapped in a proper single-voice SysEx message — which is exactly what this tool does, for all 32 voices in one go.

Where to get DX7 patches

Original DX7 patch collections (ROM cartridges and more) can be downloaded here:

https://yamahablackboxes.com/collection/yamaha-dx7-synthesizer/patches/

Grab any .syx bank from that collection, run it through dx7toFM1.py, and you get a folder of individually named voices.

Requirements

Python 3.8 or newer. Standard library only — no pip install needed.

Quick start

python3 dx7toFM1.py rom1a.syx

rom1a.syx — split: 32 voice(s)
  01  'BRASS   1'    -> BRASS 1.syx
  02  'BRASS   2'    -> BRASS 2.syx
  03  'BRASS   3'    -> BRASS 3.syx
  ...
  32  'TAKE OFF'     -> TAKE OFF.syx

Total: 32 voice(s).

Every file in rom1a_voices/ is a valid 163-byte DX7 single-voice dump that FM1's third-party firmware can import directly. The filename is the voice name stored inside the bank.

Usage

python3 dx7toFM1.py <bank.syx> [<bank2.syx> ...] [options]

Option	Description
`-o, --out DIR`	Output directory (default: `<input>_voices`)
`-n, --numbered`	Prefix filenames with a two-digit index: `01_BRASS 1.syx`
`-p, --prefix STR`	Add a fixed prefix to every filename
`-u, --underscore`	Turn spaces in voice names into `_`: `BRASS_1.syx`
`-l, --list`	Only list the voice names, write no files
`-q, --quiet`	Do not print the per-voice listing
`--selftest`	Run the pack/unpack round-trip self-test and exit

Examples:

# default: ./rom1a_voices/BRASS 1.syx ...
python3 dx7toFM1.py rom1a.syx

# numbered + underscore names into a chosen folder
python3 dx7toFM1.py rom1a.syx -o voices -n -u

# batch-convert a whole folder of banks (each gets its own subfolder)
python3 dx7toFM1.py *.syx -o voices

# just look at what is inside a bank
python3 dx7toFM1.py rom1a.syx --list

Output

Each written file is byte-for-byte a standard DX7 single-voice dump:

F0 43 0n 00 01 1B | 155 bytes of unpacked voice data | checksum | F7

The checksum is the standard Yamaha 7-bit checksum, (-sum) & 0x7F.
The device/channel number 0n is copied from the source bank.
The voice name is taken from the 10 name characters stored in the bank.

File naming rules:

Trailing padding spaces are trimmed (BRASS 1 → BRASS 1), inner spaces are kept.
Characters that are illegal in filenames (/ \ : * ? " < > |) become _.
Duplicate names are de-duplicated with a suffix: BRASS 1.syx, BRASS 1-2.syx, ….
A voice with a blank name falls back to voice_01, voice_02, ….

How it works

Bank message (4104 bytes):

F0 43 0n 09 20 00 | 4096 bytes | checksum | F7
                     32 voices × 128 bytes, packed

Packed voice (128 bytes) — 6 operators × 17 bytes, then globals, then the name:

Bytes	Contents
0–16	OP6 (operator 6)
17–33	OP5
34–50	OP4
51–67	OP3
68–84	OP2
85–101	OP1
102–109	Pitch EG rates 1–4, levels 1–4
110	Algorithm (0–31)
111	bit 3 osc key sync, bits 2–0 feedback
112–115	LFO speed, delay, pitch mod depth, amp mod depth
116	bits 6–4 pitch mod sens, 3–1 LFO wave, 0 LFO sync
117	Transpose (0–48, 24 = centre)
118–127	Voice name, 10 ASCII characters

Packed operator (17 bytes) — the crowded one, bit fields sharing a byte:

Byte	Contents
0–3	EG rates 1–4 (0–99)
4–7	EG levels 1–4 (0–99)
8	Keyboard level scaling break point
9	Left depth
10	Right depth
11	bits 3–2 right curve, bits 1–0 left curve (0–3)
12	bits 6–3 detune (0–14), bits 2–0 rate scaling (0–7)
13	bits 4–2 key velocity sens (0–7), bits 1–0 amp mod sens (0–3)
14	Operator output level (0–99)
15	bits 5–1 frequency coarse (0–31), bit 0 oscillator mode
16	Frequency fine (0–99)

Unpacked single voice (155 bytes) — one value per byte:

Bytes	Contents
0–20, 21–41, …, 105–125	OP6 … OP1, 21 bytes each
126–133	Pitch EG rates 1–4, levels 1–4
134	Algorithm
135	Feedback
136	Oscillator key sync
137–140	LFO speed, delay, pitch mod depth, amp mod depth
141	LFO sync
142	LFO waveform
143	Pitch mod sensitivity
144	Transpose
145–154	Voice name

The tool implements both directions (unpack_voice() / pack_voice()) and uses the inverse function as a self-check.

Verification

All 32 voices of rom1a.syx: pack(unpack(voice)) reproduces the original 128 bytes byte-for-byte — no bits dropped or misplaced.
Feeding tine-ep.syx (a known good single-voice dump) through the tool produces a file that is byte-identical to the original, checksum included.
Voice names recovered from rom1a.syx match the official ROM1A voice list.
python3 dx7toFM1.py --selftest runs 2000 randomised pack/unpack round-trips.

Notes and limitations

One-way only: it splits banks into single voices; it does not merge single voices back into a 32-voice bank.
Input that is already a single voice is handled too — it gets re-wrapped and renamed rather than duplicated.
A file containing several banks back to back is split completely.
Damaged checksums are reported as warnings rather than aborting the conversion.