As part of a general policy of open-sourcing our work at Idiap Research Institute, we are pleased to make available ISS: Idiap Speech Scripts.
ISS is a collection of scripts for maintenance of speech databases and dictionaries, and for training and testing of models for ASR. The scripts in turn are reliant on many other packages including HTK/HTS, Juicer and the ICSI speech tools.
- GMM (HTK/HTS) and MLP (QuickNet) training in the same framework
- Adaptation and speaker adaptive training
- Testing with hvite, hdecode and juicer
- Plays well with SGE (no knowledge of SGE required)
- Runs seamlessly either on the grid or interactively
- WFST composition with AT&T or OpenFST toolkits
- Large and growing collection of phone sets
- Automatic decision tree question sets for many languages
- Automatic decision tree threshold tuning using MDL or state count
- Features from HTK, SPTK or tracter (others trivially)
- HTK extended filenames, or not, your choice
- Database maintenance tools
- Out of vocabulary pronunciations via phonetisaurus
All the packages are on GitHub:
Exemplars for the Wall Street Journal and Aurora-4 tasks
Some other related speech processing packages are also available:
Juicer, a WFST based decoder for ASR
Tracter, a data-flow framework
SSP, a signal processing package written in python
Most releases are under a BSD licence, which is quite permissive.