ViaVoice TTS for Speech-Dispatcher

A speech-dispatcher module for IBM ViaVoice TTS 5.1. This lets you use ViaVoice as a text-to-speech engine on modern Linux through speech-dispatcher.

ViaVoice TTS is based on the same Eloquence synthesis technology that IBM and ETI developed in the 90s. Version 5.1 was freely distributed by IBM and is now archived on archive.org. It sounds significantly more natural than eSpeak NG for extended listening.

Install

Requires: speech-dispatcher, 32-bit libc support (the ViaVoice engine is a 32-bit i386 binary from 2000).

Download the latest release tarball, extract it, and run the installer:

tar -xzf viavoice-tts-bundle.tar.gz
cd viavoice-bundle

# System-wide (installs to /opt/ViaVoiceTTS)
sudo ./install.sh

# Or user-only (installs to ~/.local/ViaVoiceTTS)
./install.sh

The installer accepts --yes to skip the confirmation prompt and --prefix=PATH for a custom location.

If you don't have 32-bit support installed, on Debian/Ubuntu:

sudo dpkg --add-architecture i386 && sudo apt install libc6:i386

Test it:

spd-say -o viavoice "Hello, this is ViaVoice."
spd-say -o viavoice -y Flo "Hi, I'm Flo!"
spd-say -o viavoice -L   # list voices

Uninstall

# From the bundle directory, or from wherever you installed it
sudo ./uninstall.sh

# Or with a custom prefix
./uninstall.sh --prefix=/path/to/install

Voices

ViaVoice 5.1 ships with 8 English voices:

Voice	Name
0	Wade (adult male, default)
1	Flo (adult female)
2	Bobbie (child)
3	Male 2
4	Male 3
5	Female 2
6	Grandma (elderly female)
7	Grandpa (elderly male)

Select by name: spd-say -o viavoice -y Grandpa "Back in my day..."

Configuration

The config file lives at <install-path>/etc/viavoice.conf and is also copied to speech-dispatcher's module config directory during install. All settings are optional. The defaults work fine.

# Sample rate: 8000, 11025, or 22050 (default: 22050)
ViaVoiceSampleRate 22050

# Default voice (0-7, see table above)
ViaVoiceDefaultVoice 0

# Voice parameters (applied to all voices)
ViaVoicePitchBaseline 65     # 0-100
ViaVoicePitchFluctuation 30  # 0-100, higher = more expressive
ViaVoiceSpeed 50             # 0-250
ViaVoiceVolume 90            # 0-100
ViaVoiceHeadSize 50          # 0-100, affects resonance
ViaVoiceRoughness 0          # 0-100, voice gravel
ViaVoiceBreathiness 0        # 0-100, airy quality

# Phrase prediction (0=off, 1=on, default: off)
ViaVoicePhrasePrediction 0

# Custom dictionaries
ViaVoiceMainDict /path/to/main.dct
ViaVoiceRootDict /path/to/root.dct
ViaVoiceAbbrevDict /path/to/abbrev.dct

Building from source

Build dependencies (Debian/Ubuntu):

sudo apt install gcc-multilib libc6-dev-i386 libspeechd-dev curl rpm2cpio cpio binutils make

Build:

./scripts/build-bundle.sh

This downloads the ViaVoice RTK/SDK from archive.org, extracts them, downloads the ancient libstdc++ that ViaVoice needs, compiles the speech-dispatcher module as a 32-bit binary, and packages everything into dist/viavoice-tts-bundle.tar.gz.

All downloads are verified against embedded SHA256 checksums. Pass --skip-verify to bypass this during development.

How it works

This section explains the full pipeline from speech-dispatcher to audio output.

Architecture overview

Speech-dispatcher (SPD) is a server that sits between applications and TTS engines. Applications send text to SPD via the SSIP protocol. SPD processes the text (punctuation handling, SSML wrapping) and routes it to a module -- this project is one such module.

SPD modules are standalone executables that communicate with the SPD server over stdin/stdout using a line-based protocol. The module receives commands like SPEAK, STOP, SET, and LIST VOICES. When SPD sends text to speak, it wraps it in SSML and escapes special characters as XML entities (' becomes ', & becomes &, etc.).

The wrapper script

SPD launches sd_viavoice, which is a bash wrapper script (bundle/sd_viavoice.in). It resolves its own path (following symlinks), then sets up the environment:

• ECIINI -- points to eci.ini, the ViaVoice voice configuration file
• LD_LIBRARY_PATH -- prepends the bundle's usr/lib/ so the linker finds libibmeci50.so and the ancient libstdc++-libc6.1-1.so.2
• LD_PRELOAD -- preloads enu50.so (the English voice data) to work around a loading issue in ViaVoice

It then execs the actual binary sd_viavoice.bin.

The module binary

sd_viavoice.bin is a 32-bit ELF binary compiled from src/sd_viavoice.c and the module framework files (module_main.c, module_readline.c, module_process.c). The framework handles the stdin/stdout protocol with SPD. The module code implements these callbacks:

• module_config() -- parses the config file
• module_init() -- creates an ECI instance, allocates an audio buffer, registers the audio callback, configures voice parameters and dictionaries
• module_set() -- handles SPD parameter changes (voice, rate, pitch, volume), mapping SPD's -100..+100 ranges to ViaVoice's native ranges
• module_speak_sync() -- the main synthesis function (see below)
• module_stop() / module_pause() -- sets a flag and calls eciStop()
• module_list_voices() -- returns the 8 ViaVoice preset voices
• module_close() -- cleans up ECI handle, dictionaries, and buffers

Text processing pipeline

When module_speak_sync() receives text from SPD, it goes through these stages:

1. SSML stripping -- Removes all XML tags (<speak>, <voice>, etc.) since ViaVoice doesn't understand SSML.

2. XML entity decoding -- Converts ' back to ', & to &, < to <, > to >, " to ".

3. Text sanitization (only for normal text reading, not character-by-character or key echo):

   • Letters, digits, whitespace, basic sentence punctuation (. , ! ?), $, and ' pass through unchanged.
   • Clause-break characters (; : ( ) [ ] { }) are replaced with a comma attached to the preceding word. This preserves natural pause inflection without ViaVoice reading the character name aloud. For example, word (aside) more becomes word, aside, more. If the clause-break is at the start of text with no preceding word, it's simply dropped.
   • Punctuation immediately after a clause-break char is consumed (e.g., "). doesn't leave an isolated period).
   • Em-dashes and en-dashes (U+2014, U+2013) get the same comma treatment.
   • UTF-8 currency symbols are replaced with English words: £ to "pound", ¢ to "cent", ¥ to "yen", € to "euro". The ASCII $ passes through directly since ViaVoice handles it natively (reads "$5" as "five dollars").
   • All other characters (including multi-byte UTF-8 that ViaVoice can't handle) are replaced with spaces.

   The sanitizer allocates a new buffer (2x input length) rather than editing in-place, since clause-break expansion can produce more bytes than the input.

4. ECI synthesis -- The cleaned text is passed to eciAddText(), then eciSynthesize() and eciSynchronize(). ViaVoice runs in plain text mode (eciInputType = 0) so it applies natural prosody to punctuation (trailing off at commas, rising pitch at question marks, finality at periods) rather than reading punctuation characters aloud.

Audio path

ViaVoice synthesizes audio in chunks. An ECI callback (eci_callback) is called for each chunk with a buffer of 16-bit PCM samples. The callback appends these to a growing AudioData buffer (protected by a mutex). After synthesis completes, the full buffer is sent to the SPD server as a single AudioTrack (16-bit, mono, at the configured sample rate). SPD handles the actual audio output.

The bundle

The tarball contains everything ViaVoice needs to run:

• sd_viavoice.bin -- the 32-bit speech-dispatcher module
• libibmeci50.so -- the ViaVoice ECI engine
• enu50.so -- English voice data
• libstdc++-libc6.1-1.so.2 -- ancient libstdc++ from gcc 2.95 that ViaVoice was linked against
• ViaVoiceTTS/bin/ -- IBM SDK tools (inigen, vieweci, showmsg)
• ViaVoiceTTS/eci.ini -- voice configuration (paths are fixed up by the installer)

The only thing the host system needs to provide is 32-bit libc (libc6:i386 on Debian) and speech-dispatcher.

Why not IBMTTS 6.7 / Voxin?

IBMTTS 6.7 (sold as Voxin) is the newer version of this engine but has several issues: non-toggleable phrase prediction, non-disableable abbreviation dictionaries, and a bug where speech parameters (rate, volume) randomly reset during synthesis. ViaVoice 5.1 doesn't have these problems.

License

• Module code: BSD (based on speech-dispatcher's skeleton module by Samuel Thibault)
• ViaVoice RTK/SDK: IBM proprietary (abandonware, archived for accessibility use)

Credits

• IBM ViaVoice TTS 5.1 (~2000)
• speech-dispatcher
• Archive.org for preserving the ViaVoice packages