The first Google text-to-speech voices, which shipped only with Android 4.0 "Ice Cream Sandwich" (Google's "Flint" engine), running again on Windows: as an NVDA synthesizer and a SAPI 5 voice.

Flint is built on the HMM-based speech synthesis system HTS. This project decodes Google's compact voice format, converts it for the open-source hts_engine API, and supplies the text front end Google's engine needs: pronunciations from the CMU Pronouncing Dictionary and the NRL letter-to-sound rules, with Flint's own label format.

No Google data is included. The voices belong to Google. You supply an Android 4.0 GoogleTTS.apk, or an Ice Cream Sandwich gapps package that contains one, and the voices are converted on your own computer.

Status: early releases (0.x), in testing.

1. Install googleIcsTts-<version>.nvda-addon from the releases and restart NVDA.

2. Open NVDA menu > Tools > Google TTS (Ice Cream Sandwich) voices, then either

   • Download the voices from SourceForge: reads only GoogleTTS.apk (about 4 MB) out of the Ice Cream Sandwich gapps package gapps-ics-20120429-signed.zip, or
   • Import from a GoogleTTS.apk or gapps file: any Android 4.0 GoogleTTS.apk, or a gapps package containing it.

   Each voice is checked against the known Android 4.0 voice before it is converted. Voices are stored in your NVDA configuration folder (googleIcsTts\voices), so add-on updates keep them.

3. Choose Google TTS (Ice Cream Sandwich) in NVDA's Select Synthesizer dialog (NVDA+Ctrl+S). Two voices: English US (female) and English GB (male, RP).

Voice settings include NVDA's rate boost and Shorten pauses, which trims the silence between the pieces NVDA speaks.

Works with NVDA 2023.1 and later, 32- and 64-bit, including Windows 7 with NVDA 2023.3. Check for updates in the same dialog looks for a newer release on GitHub, only when you press it.

For JAWS, Narrator, Balabolka and any other program that speaks with SAPI 5 voices; no Python or other runtime is needed.

1. Run googleIcsTts-sapi-<version>.exe from the releases. It installs both the 64- and 32-bit voice, so 32-bit programs get it too.

2. Leave Download the voices from SourceForge now ticked at the end, or open Google TTS (Ice Cream Sandwich) voice settings from the Start menu later. The settings app can

   • download the voices from SourceForge, or import them from a GoogleTTS.apk or gapps file,
   • copy the voices the NVDA add-on already imported,
   • use a voices folder of your own (one holding ics_en_us.htsvoice / ics_en_gb.htsvoice),
   • shorten pauses (on by default), and
   • check for updates.

   Voices are stored in %ProgramData%\Google ICS TTS\voices, which every account and the Windows sign-in screen can use. The NVDA add-on finds them there as well.

3. Pick Google TTS (Ice Cream Sandwich) English US or English GB in your program's voice list.

Rate 0 is the voice's normal speed and +10 about six times as fast; bookmarks are reported where they fall in the audio. Windows 7 SP1 and later.

• src/addon/ - the NVDA add-on: driver, voice importer, update check, notices
• src/native/ - the shared C engine: hts_engine API (vendored), the text front end (ics_front.c), the voice converter (ics_convert.c) and their API (ics_hts.h); miniz is vendored under third_party/
• src/sapi/ - the SAPI 5 voice, its settings app (native importer and update check) and the installer (installer/, Inno Setup 6)
• tests/ - parity tests against the Python reference in tests/reference/
• tools/ - voice-format decoders, the Python converter, rule generator, build and development utilities
• data/ - the CMU Pronouncing Dictionary

Needs Visual Studio or its Build Tools with the C++ workload (found with vswhere), and Python 3.

src\native\build.cmd          rem builds src\addon\engine\x64 and x86\ics_hts.dll
python tools\build_addon.py   rem writes build\googleIcsTts-<version>.nvda-addon
src\sapi\build.cmd            rem build\sapi: both ics_sapi.dll, IcsTtsSettings.exe, test probes
src\sapi\build.cmd installer  rem ...and build\googleIcsTts-sapi-<version>.exe (Inno Setup 6)

Both builds take the version from src\addon\manifest.ini. To try a SAPI build without the installer, register both DLLs from an administrator prompt (System32 regsvr32 for x64, SysWOW64 for x86).

The DLLs link the C runtime statically, so nothing else needs installing, and they load on Windows 7. The add-on supports NVDA 2023.1 and later, 32- and 64-bit.

The C front end and converter are checked against the Python reference in tests/reference/:

python tests\test_front.py                        rem labels: every CMUdict word, both accents
python tests\test_convert.py <eng-usa/voice> ...  rem C and Python .htsvoice are byte-identical
python tests\test_speak.py <us.htsvoice> <gb.htsvoice>  rem audio is identical sample for sample

The registered SAPI voice is driven through ISpVoice, both architectures and both voices: audio, bookmark order and timing, and how quickly an abort stops:

python tests\test_sapi.py

The SAPI side's native importer must produce the same voices as the Python converter, and its update check must pick the right release:

python tests\test_sapi_import.py <GoogleTTS.apk> [<gapps zip>] [--download]
python tests\test_sapi_updates.py [--live]

Nothing committed may contain a path from this machine; check before pushing (--history checks every commit too):

python tools\check_paths.py
python tests\test_check_paths.py

BSD 3-Clause for this project's code (see LICENSE). hts_engine API, the CMU Pronouncing Dictionary and the public-domain NRL rules keep their own terms; see src/addon/licenses/.