Skip to content

Latest commit

 

History

History
76 lines (48 loc) · 3.01 KB

README.md

File metadata and controls

76 lines (48 loc) · 3.01 KB
Raw

SAPI-Bridge

What are we trying to do?

  • Create a SAPI engine
  • No synth directly in the dll. Pass all speak calls - infact all calls for other methods to a pipe service (see VoiceServer)

So to be clear

  • Have a pipe service running in the background which deals with all sapi calls
  • A GUI "Register your voice" app (see VoiceServer/RegisterVoice)
  • A SAPI Dll that is registered in windows which for all purposes to windows works as a sapi system but redirects to a pipe service

Why like this? Why not build all on the dll?

Because some TTS systems we are using have really slow cold-start times to call each time. Take for example sherpa-onnx models. Its far faster to hold the client object in memory and next synth streaming calls are faster

Are you sure? There are going to be loads of methods you cant support

For sure - we arent totally sure. Infact we can do lots to improve coldstart times - but we already have most of the service code working and a pretty decent tts-wrapper that allows us to do synth streaming calls across a wide range of speech engines that are on and offline.

For all the methods - you are right - we are going to struggle to do things like wordEvents for all engines. But for some we already have this (https://github.com/willwade/tts-wrapper?tab=readme-ov-file#feature-set-overview)

What was your code originally doing here

We built a simple proof of concept to build a SAPI dll and integrate it into python. Its a total PITA to build and - we figured why not just abstract this and remove the python headaches. Redirect to a pipe service..

WE HAVENT MIGRATED AWAY FROM THIS DLL CODE MUCH - INFACT IT JUST DOESNT WORK

** So what should this dll do?**

  1. Register as a SAPI engine
  2. When voices call this engine it redirects speak calls - and other callls it can support to a pipe service
  3. We should safely pass on calls we cant support

** How could/should we test this?**

Use https://www.cross-plus-a.com/balabolka.htm and see if the voices get registered (they do now) but they dont speak

So what follows below is specific code details on the DLL part..

Building

Use Visual Studio to build the project. The project is a CMake project, so you can use CMake to generate the Visual Studio solution file So the build process is as follows:

  1. Download opensource jsoncpp.
  2. Build jsoncpp.lib and jsoncpp.dll
  3. Link jsoncpp.lib and dll to pysapittsengine project.
  4. Then build the pysapittsengine.dll
cd engine
mkdir build
cd build
cmake ..
cmake --build . --config Debug
cmake --build . --config Release

Registering engine (run as Administrator)

regsvr32.exe pysapittsengine.dll

Registering voice (run as Administrator)

regvoice.exe --token PYTTS-AzureNeural --name "Azure Neural" --vendor Microsoft --path C:\Work\SAPI-POC;C:\Work\build\venv\Lib\site-packages --module voices --class AzureNeuralVoice

Or use the GUI to register voices. See VoiceServer/README.md for more information.