Skip to content
Acoustic and language models for minorised languages.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
ca-es
img
scripts
.gitattributes
.gitignore
LICENSE
README.md
requirements.txt

README.md

CMU Sphinx Models

The repository of acoustic and language models, and the tools to download raw data and train them. Currently only for Catalan.

The models are trained with sphinxtrain 5prealpha. You can find further details in our webpage and in our wiki. If you just want to download the latest acoustic and language models in Catalan you can find them here.

git clone

Once you have complete the standard git clone you also need git-lfs.

The first time you need to:

sudo apt-get install git-lfs
git lfs install

and then every time:

git lfs pull

Requirements

For basic setup, one needs to install PocketSphinx. For Debian systems:

sudo apt-get install pocketsphinx

You can also compile directly from the releases, or from the source code in github or sourceforge. The installation steps are explained in the official CMU Sphinx tutorials.

The test scripts are tested for Python 3.5.2, and in order for them to run, pocketsphinx and SpeechRecognition modules are needed. You can install them by:

sudo apt-get install swig
sudo apt-get install libpulse-dev
pip3 install -r requirements.txt

Catalan

Models

All the files necessary for pocketsphinx, i.e. the language model, phonetic dictionary and the acoustic models are in ca-es directory. The model files are:

ca-es
 ├─ pronounciation.dict       (Phonetic dictionary)
 ├─ language-model.lm.bin     (Language model)
 └─ acoustic-model            (Acoustic model)
    ├─ feat.params
    ├─ mdef
    ├─ means
    ├─ mixture_weights
    ├─ noisdict
    ├─ transition_matrices
    └─ variances

In order to test the models, simply execute:

$ python scripts/decode_from_file.py
la seva abraçada havia estat una batalla el clímax una victòria

Corpus

The training data consists of 240 hours of Catalan Public television (TV3Parla corpus) and 320 hours of recording from Catalan Parliament (Parlament de Catalunya) plenary sessions (ParlamentParla corpus). The TV3 data was prepared with the support of Softcatalà, the Parlament data was prepared with the support of the Culture Department of the Catalan autonomous government.

For further details and the download links of the data, please visit our resurces for speech technologies page.


Part of this project, specifically the ParlamentParla corpus and training of current acoustic models, was possible thanks to the support of the Culture Department of the Catalan autonomous government.

You can’t perform that action at this time.