.FST File Format

Introduction

A finite-state transducer is a graph that turns one sequence of symbols into another: the states are joined by arcs that read an input symbol, write an output symbol and carry a weight. Speech recognition and text processing use them for pronunciation lexicons and language models. The OpenFst library saves them in a binary file, usually with the extension .fst.

The format here was checked on eight files, the pmap.*.fst pronunciation maps in macOS’s SpeechRecognitionSharedSupport.framework: the header, the two symbol tables and the one state with its arcs. All eight are the vector type with standard arcs, so other types (const, compact and so on) are not described, and the page is marked partial.

The extension is shared with FL Studio presets, which start with “FLhd”, and with F* source files, which are text. The first four bytes tell them apart.

Structure

Everything is little-endian. A string is a 4-byte length and then that many bytes, with no terminator.

In the vector type each state is a final weight (4-byte float; 0 in these files), an arc count (8 bytes) and then that many 16-byte arcs: input label, output label (both 4 bytes, 0 is epsilon), weight (4-byte float) and next state (4 bytes). The eight files have one state, and every arc leads back to state 0: the transducer maps phoneme symbols to output symbols with a cost, for example the second arc of pmap.en_GB.fst is input 0, output 2, weight 4.0215.

A symbol table is the magic 74 FB B2 7E (0x7EB2FB74), the table’s name as a string, the next free key (8 bytes), the number of symbols (8 bytes), then each symbol as a string followed by its key (8 bytes). In the macOS files the name is a path from the vendor’s build, the tables have 38 to 55 symbols that begin <eps> (key 0), AA, AE, AH in the English files, and the output table has two more symbols than the input table in every file.

Worked example

00000000  d6 fd b2 7e                                      magic
00000004  06 00 00 00  76 65 63 74 6f 72                   "vector"
0000000e  08 00 00 00  73 74 61 6e 64 61 72 64             "standard"
0000001a  02 00 00 00  03 00 00 00                          version 2, flags 3
00000022  03 00 82 15 81 00 00 00                          properties
0000002a  00..(8)  01 00..(8)  00..(8)                      start 0, 1 state, 0 arcs
00000042  74 fb b2 7e  56 00 00 00  2f 61 6d 72 ...         symbol table, name of 86 bytes

Not described

This page is partial. It rests on eight files that are all the vector type with standard arcs and one state.

Also used for

• FL Studio channel preset – File type 32: one channel of the channel rack, with its sample or plugin.
• FL Studio mixer preset – File type 64: one mixer insert, with its effect plugin and its settings.
• FL Studio plugin preset – File type 48: one plugin’s name and saved state, with no channel around it.
• FL Studio preset – FL Studio channel, plugin and mixer presets: an FLP file with another file type.

Links