SpeechX Architecture

Directory Tree

The directory structure is as follows：

speechx/
├── CMakeLists.txt
├── docker    # dockerfiles
├── examples  # asr, tts examples
└── speechx   # lib .a/.so
    ├── CMakeLists.txt
    ├── decoder   # decoder
    │   └── CMakeLists.txt
    ├── frontend  # feature
    │   ├── CMakeLists.txt
    │   ├── audio
    │   │   └── CMakeLists.txt
    │   └── text
    │       └── CMakeLists.txt
    ├── kaldi    # kaldi libs
    ├── model    # inference engine
    │   └── CMakeLists.txt
    ├── protocol # http/websoket/grpc
    │   └── CMakeLists.txt
    ├── third_party
    │   └── CMakeLists.txt
    └── utils
        └── CMakeLists.txt

12 directories, 10 files

Common Modules

• gflag，glog，gtest：openfst 使用的是 gflag/glog，替换KALDI_WARN等。

• utils: absl-cpp

• config模块：保留kaldi config。

• GRPC才考虑Protobuff

• tcmalloc: 并发时性能好

• mempool: 最好能一起管理cpu和gpu的mem；减少内存分配开销和碎片

• threadpool：做单core多worker可以用。 https://github.com/progschj/ThreadPool

• queue: 需要queue在 data -> feat -> decode 之间做异步操作

• decode task: 需要支持单core多线程