Download IMS German Festival
The German Festival version was developed at the Institute of Natural Language Processing, University of Stuttgart by Gregor Möhler, Antje Schweitzer, Mark Breitenbücher, Martin Barbisch, and others.
IMS German Festival is a German Text-to-Speech system which we provide as an extension to the original Festival distribution by CSTR, University of Edinburgh. Together with two non-IMS resources (the BOMP lexicon from the University of Bonn, which is included in the distribution, and diphones from the MBROLA project) it enables Festival to speak German.
IMS German Festival contains all modules necessary for text-to-speech conversion. More explicitly this involves linguistic text analysis (modules tokenization, token-pos, token-to-word), grapheme-to-phoneme conversion (lexicon-lookup, letter-to-sound rules), prosody (phrasing, intonation, duration, intonation targets) and waveform synthesis (carried out by the MBROLA engine through an appropriate interface).
For more background information about the IMS German Festival see the README file and the COPYING file of the distribution or consult the manual included in the distribution.
The packages are distributed for research and educational (i.e. non-commercial and non-military) purposes. The license terms can be found in the file COPYING.ims_german_festival for the IMS German Festival package and in file COPYING.bomp for the BOMP lexicon.
NEW!! This is a new (beta) Version, IMS German Festival 1.3-os(beta). It is compatible with Festival_2.1 as well as with Festival_1.95 and Festival_1.96. The IMS German Festival 1.2 sources are also provided in a separate directory on the download page, in case you prefer them.
PLEASE NOTE. There was an error in the recent distribution which can cause the following error message when loading voices:
SIOD ERROR: not a symbol or string : #<CLOSURE (utt) (begin "(MBROLA_Synth UTT)
Synthesize using MBROLA as external module. Basically dump the info
from this utterance. Call MBROLA and reload the waveform into utt.
[see MBROLA]" (let ((filename (make_tmp_filename))) (save_segments_mbrola utt filename) (system (string-append mbrola_progname " " mbrola_database " " filename " " filename ".au")) (utt.import.wave utt (string-append filename ".au")) (delete-file filename) (delete-file (string-append filename ".au")) utt))>
To fix this, please exchange this file in festival/lib/german/: ims_german_voices_opensource.scm, or download the patch ims_german_1.3-os.fix.tgz provided on the download page.
These are the changes compared to version IMS German Festival 1.2-os:
- compatible with Festival 2.1
- handles short, tense vowels
- no error when no voiced segment in syllable
- supports MBROLA voice de4, with glottalized and unglottalized vowels
- can deal with British English 9: in BOMP lexicon
- improved duration for nasals
- new text preprocessing, hopefully improved! but still far from perfect (numbers, abbreviations, etc.)
- can theoretically deal with French accents in text (but reasonable pronunciation only if words are added in lexicon!)
- less unknown_diphones warnings
- better durations, and also pause lengths reduced, especially utterance-initial and utt-final
- more reasonable voice resets
- BOMP lexicon improved: distinguish between A~ and o~, some errors corrected
- festival/examples/text2wave should work with German voices now
- Festival 2.1 (speech_tools and festival itself)
compiled on a Linux machine (with GCC)
- MBROLA engine and at least one German MBROLA voice
from the MBROLA project page
To get the distribution please fill out this form and you will receive an email with the address of the hidden download-site. This data is only used to keep an overview over the community using the IMS German Festival and is not distributed in any way.
Please direct inquiries to german-festival AT ims.uni-stuttgart.de.
We provide a demo where you can synthesize any given text online. The demo provides diphone synthesis with synthesis modules developed at IMS. Not all of these are contained in our open source version; some modules in the open source version are simpler than these, so the synthesis results from the open source version may sound different from what you get from the open source version.