We are pleased to announce that CMUSphinx project is accepted to Google Summer Of Code 2012 program. That will enable us to help several students to start their way in speech recognition, open source development and in CMUSphinx. We are really excited about that.
If you are interested to participate as a student, an application period will open soon but it’s better to start preparation of your application right now. Feel free to contact us for any questoins! For more details see:
If you would like to be a mentor please sign in into gsoc web application and add your ideas to the ideas list:
We invite you to participate!
The iSound is a program that was built with help of the CMU Sphinx-4 system. It is a part of the thesis at the Faculty of Mathematics, Natural Sciences and Information Technologies from Koper, Slovenia. Its main goal is real-time audio signal visualization, also known as spectrogram or sonogram. Which means, that it allows observation of the sound.
This property could be useful in many areas, such as: phonetics, animal sounds analysis, music, sonar/radar, speech processing, seismology, etc. Additionally, it has included few features into basic spectrogram drawing, which made the application more useful. That features are: image freezing, zoom control, signal frequency display, resizing, changing of the color schemes and contrast adjustment.
Compared to the other programs with similar functioning it gives promising results of the CPU, memory and graphics usage. Tests were made on Windows XP, Windows 7, Linux ubuntu and OS X Lion. You can find full test results in the diploma, on page 40.
Author’s comment: “For future work i plan to publish research work as an article in the journal. Currently I’m working on idea, how to use similar technologies and develop a tool, which can help persons with hearing handicap.”
Find useful information about the project, at the author, Irman Abdić’s web page:
In many languages the amount of lexical forms is huge due to morphology. Even simple vocabulary can contain several million forms and variations. It's hard to recognize such a big vocabulary because of huge search space. Decoder is slow and a language model takes enormous amount of memory.
Of course brute force approach make sense and actually quite successful but better ones already suggested. For example using morphological segmener we can build a language model and the acoustic model which can describe the same vocabulary in way smaller number of subword items. Real words are combined from the chunks which are separate entities in a language model. This way our search space is efficiently represented and the speed is comparable to English models.
The tricky part is to properly segment the words. Because pronunciaiton of decomposition is not so straightforward it takes some effort to build the split. We are happy that our contributor Zamir Ostroukhov managed to solve that problem. He created the acoustic model from the audiobooks from the Voxforge database and used large text corpora to create a morphologically-segmented language model. This is a very promising approach for morphologically-rich language so we look forward to see this framework as a part of CMUSphinx. Maybe this framework could be extended to multilevel speech representation which could hold both subwords and sentence-level items.
Check Zamir's project
For more details on the approach please see
Large vocabulary continuous speech recognition of an inflected language using stems and endings by Toma Rotovnik at al.
Download Russian audiobook model here, the morphological language model is included:
http://sourceforge.net/projects/cmusphinx/files/Acoustic and Language Models/Russian Audiobook Morphology Zero
For more details see
Over the last couple of weeks Long Audio Alignment Project has had a lot of new developments. It was understood that the accuracy of audio alignment would improve even further if some approximate time information for certain words were known from before the actual alignment. A decoder without such an information, during alignment has to go through all the frames of the audio to finally tell which alignment hypothesis scores the best. However, with such approximate timed information the decoder only has to wait until one (or more) of it's hypothesis token agree with this additional information, allowing it to prune out the rest of the tokens. This helps keeping the beam size small.
A highly configurable phrase spotter was hence implemented to get this timed information. Phrase spotter creates a left to right no skips grammar from the words in the phrase and (like Aligner) uses a garbage model to model all out of grammar utterances. The grammar has been chosen this way to ensure that a phrase is recognized only when all the words in the phrase are present in the utterance without a skip. Corresponding changes in the linguist were made as well to ( allow and ) ensure that one a Phone Loop is inserted only at the start of a phrase in the search graph.
Aligner search manager was then designed to exploit this phrase spotter's result and perform audio alignment. As a result, even with much more complicated grammar, the aligner can now align audio with much better accuracy (almost 0% error) , however memory requirements for aligning very large text allowing large skips still poses a problem. For now I am profiling the memory usage to locate and tackle this issue.