apple-eloquence-elf

Convert Apple's bundled ETI Eloquence TTS engine (Mach-O dylibs from the TextToSpeechKona framework) to Linux ELF shared objects, and ship them behind a native speech-dispatcher module.

Important

AI-assisted development. Most of this project was implemented in pair-programming sessions with Anthropic's Claude. Every change is human-reviewed and validated through end-to-end speech testing with Orca + speech-dispatcher and container-based install testing on Arch, Debian, Ubuntu, and Fedora — but the implementation work itself is largely AI-generated. The Co-Authored-By trailer on each commit makes the split visible. Decide whether you're comfortable with that posture before depending on it.

$ ./examples/speak ./prebuilt/x86_64/eci.so "Hello world."
eciVersion: '6.1.0.0'
PCM: 27907 samples (2.53s @ 11025Hz), peak amplitude 18103
Wrote /tmp/eci_out.s16
$ aplay -r 11025 -f S16_LE /tmp/eci_out.s16

What this is

Apple ships ETI Eloquence as part of VoiceOver across macOS, iOS, iPadOS, and tvOS. Inside TextToSpeechKonaSupport.framework they bundle dylibs that are the ETI ECI 6.1 engine compiled for Apple platforms. The dylibs depend only on libSystem.B.dylib and libc++.1.dylib, which makes them tractable to retarget.

Eloquence originated at Eloquent Technologies, Inc. (ETI). SpeechWorks acquired ETI in 2000; ScanSoft acquired SpeechWorks in 2003 and renamed itself Nuance Communications in 2005. Nuance spun off its automotive AI division as Cerence in 2019, taking the Eloquence / Vocalizer TTS stacks with it. Microsoft's 2022 acquisition of Nuance did not include those TTS engines, so Eloquence's current owner is Cerence, not Microsoft. IBM had its own ECI-licensed fork shipped as ViaVoice / IBMTTS; the mainline engine — including what Apple ships — descends from the ETI tree.

macho2elf.py is the converter: a Python + LIEF tool that produces ELF .so files exposing the standard ECI C API.

Project status

Architecture	Status
x86_64 Linux	✅ tested end-to-end with sd_eloquence + Orca
aarch64 Linux	✅ synthesis verified; output is byte-near-identical to x86_64

Shipped binaries are built from the tvOS 18.2 Simulator Runtime, for both x86_64 and aarch64. The arm64 converter gaps that previously made synthesis segfault are fixed: high8-tagged-pointer rebases, the Apple→AAPCS64 variadic ABI bridge, __chkstk_darwin, and the libc++ DT_NEEDED chain. arm64 output synthesizes the same waveform as x86_64 (differences are sub-LSB FP rounding).

Working languages on x86_64: en-US, en-GB, es-ES, es-MX, fr-FR, fr-CA, de-DE, it-IT, pt-BR, fi-FI. CJK (ja-JP, ko-KR, zh-CN, zh-TW) is gated for v1 — the romanizer init path needs the modern 2-suffixed ECI API rather than the legacy one v1 uses. Deferred to v2; background in docs/cjk-investigation/ and docs/eci-2-api/.

Install (from the release tarball)

tar -xzf apple-eloquence-elf-*-linux-x86_64.tar.gz
cd apple-eloquence-elf-*-linux-x86_64
sudo ./install.sh
spd-say -o eloquence "Hello from Eloquence."

install.sh detects the host distro family (Debian / Ubuntu / Fedora / Arch / openSUSE) and uses the system package manager to install every runtime dependency (speech-dispatcher, libc++, libsoxr, libxml2, pcre2) before dropping files into /usr/lib/eloquence/, the speech-dispatcher modulebindir, and /etc/speech-dispatcher/modules/eloquence.conf.

Configure the module by editing /etc/speech-dispatcher/modules/eloquence.conf and restarting speech-dispatcher. Audio previews of every libsoxr resampler preset ship at /usr/share/eloquence/resampler-previews/.

Uninstall: sudo ./uninstall.sh (add --purge to also remove the conf template).

Build from source

cmake -B build -DCMAKE_INSTALL_PREFIX=/usr
cmake --build build
sudo cmake --install build

cmake --install does not touch the system package manager. Install build and runtime deps yourself: cmake, gcc, llvm (for llvm-lipo), python3 with the lief package, plus dev packages for speech-dispatcher, libc++ / libc++abi, libsoxr, libxml2, and libpcre2-8. CMake will name any missing ones at configure time.

Files placed:

• /usr/bin/macho2elf — converter CLI
• <speechd modulebindir>/sd_eloquence — module binary
• /etc/speech-dispatcher/modules/eloquence.conf — config template
• /usr/share/doc/eloquence/ — README + docs

Convert your own dylibs

python3 -m venv venv && ./venv/bin/pip install lief
llvm-lipo -extract x86_64 vendor/tvOS-18.2/eci.dylib -output /tmp/eci.x86_64
./venv/bin/python3 macho2elf/macho2elf.py /tmp/eci.x86_64 -o /tmp/eci.so

Full recipe (including extraction from a tvOS Simulator Runtime DMG): docs/01-extraction.md and docs/02-conversion.md.

Speech-dispatcher module

sd_eloquence/ is a native speech-dispatcher output module. It exposes the eight Apple voice presets (Reed, Shelley, Sandy, Rocko, Flo, Grandma, Grandpa, Eddy; Jacques replaces Reed in French), transcribed verbatim from KonaVoicePresets.plist. Every supported language is selectable on the fly via speech-dispatcher's language= parameter; no per-language conf edits.

Configuration reference: docs/03-integration.md plus the comments in eloquence.conf itself.

How it works

The converter walks each section of the Mach-O dylib, emits an assembly stub that interleaves .incbin of the original code/data with section-relative labels for exports and .quad references for chained-fixup bindings, generates a linker script that pins each section at its original Mach-O virtual address (preserving RIP-relative offsets), and runs gcc -shared. Symbol prefixes are stripped (_atoi → atoi), Darwin-specific symbols are renamed (___error → __errno_location, ___stderrp → stderr), and small stub C functions cover names with no direct Linux equivalent (__maskrune, __stack_chk_guard, _DefaultRuneLocale).

Details: docs/04-internals.md.

Repo layout

macho2elf/macho2elf.py    Python + LIEF converter (MIT)
sd_eloquence/             speech-dispatcher module (GPL-2.0-or-later)
vendor/tvOS-18.2/         Unmodified Apple Mach-O dylibs
prebuilt/{x86_64,aarch64} Converted ELFs (gitignored)
examples/                 dlopen-based TTS sample (speak.c)
dist/                     install.sh / uninstall.sh / smoke.sh
tools/                    Converter audit + checksum tooling
docs/                     Extraction / conversion / integration / internals / troubleshooting
.github/workflows/        CI: convert vendor/ -> ELFs, package tarballs

Licensing

• macho2elf/ is MIT. See LICENSE.
• sd_eloquence/ is GPL-2.0-or-later. It incorporates anti-crash regex tables and dictionary-loading patterns from the NVDA-IBMTTS-Driver (Copyright (C) 2009-2026 David CM, GPL-2.0). Full text in sd_eloquence/LICENSE.GPL.
• vendor/tvOS-18.2/ are unmodified Apple binaries from the tvOS 18.2 Simulator Runtime IPSW, subject to Apple's SDK terms. SHA256 checksums in tools/checksums.txt.
• Converted .so files are derivative works of Apple's distribution. Not committed; built locally or downloaded from the release tarball.

To redistribute or productize, the conservative path is to download your own tvOS Simulator Runtime from Apple's developer portal and convert locally; see docs/01-extraction.md.

Acknowledgements

• ETI / SpeechWorks / ScanSoft / Nuance / Cerence — successive custodians of the engine.
• IBM ViaVoice TTS / IBMTTS, whose contributions to the mainline ECI codebase persist in this build.
• LevelStar, for keeping a working Linux ECI distribution alive in their Icon product; their eci.ini informed our minimal template.
• Agner Fog (objconv) and the LIEF project, for the binary-format tooling that made this tractable.
• Anthropic's Claude Code did most of the engineering pair-programming.