Muse Speech Technology Research Platform
Free download available for: Linux, BSD, Mac OS X (Apache 2.0 license) - Windows compatibility update coming soon. Download link is in the Installation section below.
Please report any issues, suggestions or contributions to peter.cahill@ucd.ie.
What is Muse?
Muse is a speech technology platform designed to be a rapid prototyping platform for speech technology research (speech synthesis and speech recognition). In order to achieve this Muse can be considered as a middle layer that connects the user with speech data, metadata, and tools. Muse aims to support a wide range of file I/O to ease using pre-recorded speech data.
The current version of muse can convert between phonetic annotation file formats (MLF, TIMIT, EmuLab, CTN, TRN, praat, sphinx).
It can also be used to filter audio/corpora, current filters wrap around HTK (MFCC, LPC), and Praat (pitch, intensity). New filters can be added easily via scripts. Some post-filtering processes are also available: Linear interpolation, Spline interpolation, first order differential coefficients (delta), second order differential coefficients (acceleration), natural log, normalisation.
Support for processing different speech corpora is also available, currently we have support for TIMIT, Switchboard, Romanian BABEL, Festvox (including CMU ARCTIC), German Kiel. Support for some of the OGI multilanguage corpora will be added in the near future.
Muse is intended to assist research into text to speech synthesis (TTS) and speech recognition (ASR).
Links
- Contains details and descriptions of the design of Muse and how to use it.
- Contains details on the functionality contained within the core Muse library.
Installation
Download the current version (last update: Sunday, 12 December 2010) (compiled binary package - suitable for most users) here:
Installation instructions are included in Chapter 1 of the Muse guide:
Download the current source version (last update: Wednesday, 13 October 2010) (source code - only required for adding new core datatypes) here:
NOTE:
This is the final release that will support Mono 2.6.7 or lower. Mono 2.8 has several significant API changes that break compatibility with this version. Muse 0.7 (for Mono 2.8+) will be released in July. This update will only affect the museshell program, none of the Muse API needs changes.
Road Map
Features for version 0.7 (TARGET: July 2011)
- Weighted finite state tools to be included (may use OpenFST rather than internal tools)
- Support for Mono 2.8
- Synthesis algorithms to be included and documented
- Pronunciation estimation algorithms to be included and documented
- Grapheme to phoneme scripts
- Speech recognition (from trained acoustic models) scripts
- Improved Windows support
Features for version 0.6 (Released: December 12)
- Several bug fixes
- Updated installer
- Improved error reporting
Features for version 0.5 (Released: July 21)
- MultiCore / Cluster support: All AI/machine learning training can utilise multicore architectures
- Support for subdirectories in the lib folder, with one subdirectory for the muse version to avoid upgrade conflicts
- Autocompletion in the muse shell
- Improved support for HTK