SFE Format

This page describes the SFE (SF-Enhanced 4) sound bank format exactly as spessasynth_core reads and writes it. SFE is a superset of SoundFont 2.04. Everything not listed here works as in the SoundFont 2.04 specification.

spessasynth_core implements a subset of the "release draft" of mine (What SFe 4.0 should have been).

Terms like BYTE, WORD, SHORT, DWORD, FourCC and chunk mean the same as in the SF2 specification. All values are little-endian.

Reading and writing

• SoundBankLoader.fromArrayBuffer loads SFE files like any other bank. The loaded bank's BasicSoundBank.type is "sfe".
• BasicSoundBank.writeSFE writes one, with SFEWriteOptions.
• Any bank can be written as SFE, no matter which format it was loaded from.

const bank = SoundBankLoader.fromArrayBuffer(input);
const sfe = bank.writeSFE({ rf64: true });
Copy

Table summary

DMOD (default modulators) and xdta (extended limits) are SF2 extensions, so they are not described here. The SFE writer always enables both.

File structure

RIFS (or RIFF)
    "sfen"
    LIST INFO
        ifil, isng, INAM, irom, iver, ICRD, IENG, IPRD, ICOP, ICMT, ISFT
        DMOD            (optional, default modulators)
        LIST xdta       (optional, extended limits)
    LIST sdta
        smpl
    LIST pdta
        phdr, pbag, pmod, pgen, inst, ibag, imod, igen, shdr
Copy

The order and layout of all standard chunks are the same as in SF2.

The main chunk

• The main chunk type is sfen instead of sfbk. This is the one deliberate break from SF2, so legacy SF2 readers refuse the file instead of playing it incorrectly.
• The reader also accepts sfbk and sfpk (SF2Pack) here, as SF2 banks.

64-bit chunk sizes (RIFS)

RIFF chunk sizes are 32-bit, which limits a file to 4 GB. When the top-level FourCC is RIFS instead of RIFF:

• Every chunk in the file (the main chunk, all LIST chunks and all sub-chunks) has a 12-byte header: the 4-byte FourCC and an 8-byte (QWORD) size.
• All other RIFF rules stay the same. The size does not include the header or the pad byte, and odd-sized chunks are followed by one pad byte. LIST chunks start their data with the 4-byte list type.

The writer uses RIFS when the rf64 option is true (the default). With rf64: false it writes a regular RIFF file with the sfen type. The reader detects the mode from the first FourCC, for any bank type.

Preset headers: the bank LSB

In SFE, the wBank field of phdr holds three values:

So an XG bank like MSB 0, LSB 66 is stored as it is, and a drum kit keeps its own MSB and LSB. These map to BasicPreset.bankMSB, BasicPreset.bankLSB and BasicPreset.isGMGSDrum.

For comparison, the SF2 writer keeps the SF2 conventions: wBank is the bank MSB, 128 for a drum preset, or the bank LSB when the MSB is 0. The reader decodes wBank with the SFE layout above for every bank, which gives the same result for regular SF2 values below 128.

Generators 61 to 88

The SF2 specification defines generators up to 60 (endOper). SFE adds generators above it, in two groups:

• 61 to 67 extend the LFOs. The GS/XG controller matrix uses them at runtime as well.
• 68 to 88 add voice features that SF2 cannot express, for example from XG voices.

Both groups follow the SF2 rules for value generators, with these additions:

• They are only stored in SFE banks. The SFE writer keeps them in pgen and igen, along with every modulator (in zones and in DMOD) whose destination is one of them.
• ignoreNoteOff, vibLfoWaveform and modLfoWaveform are only valid in instrument zones.

They are all listed in Generator Types and in GeneratorTypes, with their limits in GeneratorLimits.

LFO generators 61 to 67

With generators 63 to 65, the vibrato LFO can drive pitch, filter and volume, like the modulation LFO in SF2. GS and XG System Exclusive messages (the controller matrix) are turned into modulators targeting these generators, so a bank can use them in the same way.

Voice generators 68 to 88

Ignore note off (68)

A note off does not release the voice, and the sustain pedal does not hold it. The voice plays until its volume envelope or its sample ends, even for very short notes. This is the XG drum "Rcv Note Off" switch. A looped sample with an audible sustain plays until the channel's notes are stopped. All Notes Off, All Sound Off and stopping all voices still release it.

Delay voice (69)

The whole voice starts this long after the note on: the sample, both envelopes, both LFOs and portamento. It is silent before that. The start is rounded up to the next rendering quantum, so it can be up to 128 samples late. A note off before the voice starts does not cancel it. It starts, plays the minimum note length, then releases.

LFO fade (70, 71)

After the LFO delay (delayVibLFO or delayModLFO) ends, the LFO's output ramps linearly from 0 to full over this time. It scales all of that LFO's depths.

LFO waveform and start phase (72 to 75)

• Waveform 0 is the SF2 triangle. Waveform 1 is a rising sawtooth. Waveform 2 is a sine. All cross zero going up at the default start phase, so the sawtooth starts like the triangle and drops from +1 to -1 once per cycle.
• The start phase is added to where the LFO starts, in degrees of a cycle. 0 is the SF2 start (zero, rising), 90 starts at the peak and 180 at zero, falling.

Maximum filter cutoff (76)

This sets a ceiling for the final cutoff: initialFilterFc plus its modulators, both LFOs and both modulation envelopes. It is applied before the filter is computed, so velocity or an envelope cannot open the filter past it. The default 13500 means no ceiling.

Second volume decay (77, 78)

After the decay reaches sustainVolEnv, the volume envelope continues to sustainVolEnv + sustain2VolEnv (clamped to 0 to 960 centibels) as a straight line in centibels.

• decay2VolEnv follows the SF2 decay convention: it is the time for a full 100 dB, and shorter distances take proportionally less. keyNumToVolEnvDecay scales it like the first decay.
• A negative sustain2VolEnv makes the level rise back up.
• A release during this segment starts from the current level.
• With sustain2VolEnv at 0 there is no second segment.

Second modulation envelope (79 to 88)

This is a second envelope that works exactly like the SF2 modulation envelope (convex attack, linear decay and release, key scaling around key 60), but it is fully independent. Generators 87 and 88 add it to the pitch and the filter cutoff, like modEnvToPitch and modEnvToFilterFc. This allows, for example, separate pitch and filter envelopes. With generators 87 and 88 at 0, the envelope is not computed.

Like the first modulation envelope, it is not a modulator source. It only reaches the pitch and the cutoff, through generators 87 and 88. This keeps it exactly as the standard modulation envelope, so a synthesizer can reuse its existing envelope code for it.