SoundSwallower: an even smaller speech recognizer

"Time and change have a voice; eternity is silent. The human ear is always searching for one or the other."
Leena Krohn, Datura, or a delusion we all see

SoundSwallower is a very small and simple speech recognizer intended primarily for embedding in web applications. The goal is not to provide a fast implementation of large-vocabulary continuous speech recognition, but rather to provide a small implementation of simple, useful speech technologies.

With that in mind the current version is limited to finite-state grammar recognition. In addition, the eternally problematic and badly-designed audio library as well as (almost) all other external dependencies have been removed.

Compiling SoundSwallower

Currently SoundSwallower can be built in several different ways. To build the C library, run CMake in the standard way:

cmake -S . -B build
cmake --build build
cmake --build build --target check
sudo cmake --build --target install

Note that this isn't terribly useful as there is no command-line frontend, and shared libraries are not built by default (pass -DBUILD_SHARED_LIBS=ON if you insist). You probably want to target JavaScript or Python.

Installing the Python module and CLI

The SoundSwallower command-line is a Python module (soundswallower.cli) and can be installed using pip. It is highly recommended to do this in a virtualenv. You can simply install it from PyPI:

pip install soundswallower

Or compile from source:

pip install .

For development, you can install it in-place, but please make sure to remove any existing global installation:

pip uninstall soundswallower
pip install -e .

The command-line supports JSGF grammars and word-level force alignment for one or more input files, for example:

soundswallower --align tests/data/goforward.txt tests/data/goforward.wav
soundswallower --align-text "go forward ten meters" tests/data/goforward.wav
soundswallower --grammar tests/data/goforward.gram tests/data/goforward.wav

Note that multiple input files are not particularly useful for --align or --align-text as they will simply (try to) align the same text to each file. The output results (a list of time-aligned words) can be written to a JSON file with --output. To obtain phoneme-level alignments, add the --phone-align flag. The JSON format (which has recently changed) is the same as used in PocketSphinx 5.0 and is more compact than it is readable, but briefly, it consists of one dictionary (or "object" in JavaScript-ese) per line, where the t attribute is the recognized text and the w attribute contains a list of word segmentations, with start time in b and duration in d and, optionally, a list of phone segmentations in the w attribute with the same format.

See also the full documentation of the Python API.

Compiling to JavaScript/WebAssembly

To use the JavaScript library in your projects:

npm install soundswallower

To build the JavaScript library, use CMake with Emscripten:

emcmake cmake -S . -B jsbuild
cmake --build jsbuild

This will create soundswallower.js and soundswallower.wasm in the jsbuild directory, which you can then include in your projects. You can also use npm link to link it to your node_modules folder for development Demo applications can be seen at https://github.com/dhdaines/alignment-demo and https://github.com/dhdaines/soundswallower-demo.

To run the JavaScript tests:

cd jsbuild
npm install
npm test
npx tsc
node test_typescript.js

And in the browser:

cd jsbuild
python server.py
# Navigate to http://localhost:8000/test_web.html

For more details on the JavaScript implementation and API, see js/README.js.

See also the documentation of the JavaScript API.

Creating binary distributions for Python

To build the Python extension, I suggest using build, as it will ensure that everything is done in a totally clean environment. Run this from the top-level directory

python -m build

In all cases the resulting binary wheel (found in dist) is self-contained and should not need any other components aside from the system libraries. To create wheels that are compatible with multiple Linux distributions, see the instructions in README.manylinux.md.

Compiling on Windows in Visual Studio Code

The method for building distributions noted above will also work on Windows, from within a Conda environment, provided you have Visual Studio or the Visual Studio Build Tools installed. This is somewhat magic.

If you don't have Conda, then what you will need to do is:

• Install Visual Studio build tools. Unfortunately, a direct link does not seem to exist, but you can find them under Microsoft's downloads page. The 2019 version is probably the optimal one to use as it is compatible with all recent versions of Windows.

• Install the version of Python you wish to use.

• Launch the Visual Studio command-line prompt from the Start menu. Note that if your Python is 64-bit (recommended), you must be sure to launch the "x64 Native Command Line Prompt".

• Create and activate a virtual environment using your Python binary, which may or may not be in your AppData directory:

    %USERPROFILE%\AppData\Local\Programs\Python\Python310\python -m venv py310
    py310\scripts\activate
  

• now you can build wheels with pip, using the same method mentioned above.

Authors

SoundSwallower is based on PocketSphinx, which is based on Sphinx-II, which is based on Sphinx, which is based on Harpy, and so on, and so on, back to somewhere around the Unix Epoch. Thanks to Kevin Lenzo for releasing CMU Sphinx under a BSD license and making this possible, and Ravishankar Mosur who actually wrote most of the decoder. Many others also contributed along the way, take a look at the AUTHORS file in PocketSphinx for an idea.

This document and SoundSwallower are now being developed by David Huggins-Daines.