README.md

Speech recognition

Installation and usage

This is a Python 3 application that requires the sounddevice package to grab live frames from a mic. Install it with:

pip install sounddevice

Depending on the selected backend, additional dependencies might be required (see below).

Launch the program with --help to see available options. You can display and select the preferred input device with --list-devices and --device, respectively (otherwise the system default will be chosen).

This application opens two YARP ports: an <prefix>/rpc:s port that allows to request a dictionary/model change and to mute/unmute the microphone, and a <prefix>/result:o port that broadcasts the transcribed text. The default prefix is /speechRecognition, but it can be changed with the --prefix option.

PocketSphinx backend

Install the pocketsphinx package with:

pip install pocketsphinx

Then, launch the program with the --backend pocketsphinx --dictionary xxx --language xxx options. The dictionary and language combo relies on the adequate dictionary and model files being installed (check share/speechRecognition). For example, to use the waiter Spanish orders dictionary, put:

speechRecognition --backend pocketsphinx --dictionary waiter --language es

Vosk (Kaldi) backend

Install the vosk package with:

pip install vosk

Then, launch the program with the --backend vosk --model xxx options. Model files are downloaded on demand from the Vosk website. For example, to use the ~50 MB Spanish model, put:

speechRecognition --backend vosk --model small-es-0.42

To list and download the desired models offline and test the Vosk engine, you can use the vosk-transcriber application.