ssmc

ssmc (Speech Synthesis Markup Compiler) turns SSML into SSMIR, a three-stage intermediate representation: IR, events, and a speech plan.

SSML  →  Preprocessor  →  SSMIR ir
ir    →  Interpreter   →  SSMIR events
events → Compiler      →  SSMIR plan

The version number is only in ssmc/VERSION. The shipped tree is build/dist/ssmc-<version>/{ssmc,LICENSE,locales/}. Locale sources live in ssmc/share/locales/; see NOTICE.

Requirements

Podman or Docker. The build script installs a pinned Zig inside a throwaway Alpine image. The host does not need Zig, g++, Flex, or Bison.

Build

./build          # musl-static CLI + goldens
./build test     # compile and run goldens

Output: ../build/dist/ssmc-<version>/ (statically linked ssmc plus a copy of share/). Run from that directory so ./locales resolves.

ssmc --version prints the same number as VERSION. Runtime locales are ./locales, --locale=DIR, or SSMC_LOCALE.

Usage

ssmc file.ssml                      # SSMIR plan (default)
ssmc --emit=ir file.ssml            # preprocessor IR
ssmc --emit=events file.ssml        # interpreter events
ssmc --emit=plan file.ssml          # compiled speech plan
ssmc --locale=locales file.ssml     # CLDR-shaped locale directory
ssmc --version

--locale (or SSMC_LOCALE) is a locales root (locales/) or a language dir (locales/en). xml:lang / en-US selects en.

SSMIR 1.0

Every dump is a JSON object with a shared envelope:

{
    "ssmir": "1.0",
    "stage": "ir",
    "lang": "en-US"
}

stage is ir, events, or plan.

ir keeps the original func / text / ops model: tags are func entries with integer ids and an attrs object; each text span lists the funcs that apply (ops, innermost first). Empty val marks a <break>. <?xml?>, comments, and a wrapping <ssml> element are dropped.

events are those spans after say-as, sub, and phoneme are applied. ops stay for traceability.

plan is a linear, backend-neutral sequence of silence (ms) and speak (text, plus optional emphasis, rate, pitch, volume, ph).

Interpreter

Tag	Behavior
break	time (1s, 500ms) or strength → silence
emphasis / prosody	attach level / rate / pitch / volume to following text
say-as	W3C Note types. Numbers/dates/times are spelled with CLDR-shaped RBNF + month files we author in share/locales/<lang>/ (not a CMUdict dump, not official CLDR). Unknown interpret-as leaves the text unchanged. Time uses RBNF cardinals plus %%am / %%pm / %%oh / %%oclock particles.
phoneme	copy ph (and alphabet) onto the event
sub	replace text with alias
p / s / speak / voice	nesting / language only

There is no CMUdict lookup. Pronunciation comes from <phoneme ph="…">.

Example

tests/goldens/sample.ssml compiles to a plan with 1s / 2s silences, emphasized and slow spans, a spoken date and time, and IPA from <phoneme>.