What is TLK?
The transLectures-UPV toolkit (TLK) is an open source (Apache Licence 2.0) toolkit for automatic speech recognition (ASR) developed at Universitat Politècnica de València by the Machine Learning and Language Processing research group (MLLP).
What are the main features of TLK?
Among other functionalities, TLK features: parameter estimation of hidden Markov models (HMMs); acoustic model training; acoustic model adaptation (MLLR, CMLLR) based on deep neural networks (DNNs); and recognition (speech, text…) based on hybrid DNN-HMMs. You can find here a more detailed list of the main features of TLK.
What are the conditions to use TLK?
We would ask users to include a proper citation for TLK in all publications and products in which it is used:
The transLectures-UPV Team. TLK: The transLectures-UPV Toolkit. http://www.mllp.upv.es/tlk/
Where can I download TLK from?
Here at the TLK website.
Which operating systems does TLK support?
The current version of TLK (1.3.1) runs on Linux and Mac OS X.
How can I install TLK?
You will find instructions in the TLK website on how to install the software using pre-made packages for Mac OS X, Ubuntu and Debian, or compiling from the source code in other GNU/Linux distributions.
Which languages can TLK be used for?
TLK can be used for speech recognition (and text recognition) of any language. In order to perform speech recognition tasks, you will need statistical acoustic and language models that are fit for the language that you want to recognize.
Can I try TLK if I don’t have any statistical models?
In the TLK documentation web page, you will find three tutorials that teach different uses of TLK, which come with the data necessary to follow them.
In particular, in our tutorial on the use of tLtranscribe we provide a basic speech recognition (ASR) system for Spanish, based on freely available language resources. Using this system you can try performing speech recognition with TLK on Spanish-language videos (or audio). Please bear in mind that the Spanish ASR system provided is only a basic one (as only freely available language resources were used to train it), which will usually produce modest results; in general, the quality of the results when performing speech recognition will depend on the quality of the statistical models used (and their adaptation to the characteristics of the task).
How can I use TLK if I don’t have the statistical models for a language?
There are several options you can consider:
a) You can train statistical acoustic and language models from appropriate speech and text corpora (e.g. VoxForge, for speech, or this list of text corpora). Acoustic models can be trained using TLK (see the documentation and tutorials for more information); for language models, see for instance this list of language modelling software (TLK supports standard ARPA language models).
b) You can try the MLLP’s transcription and translation platform for speech recognition of several languages (English, Spanish, Catalan, French…; see full list). Our platform is based on TLK for speech recognition and uses the MLLP’s own advanced statistical models. Just register and start uploading your own videos or audio to see them subtitled.
How can I learn more about TLK?
You will find more detailed information in TLK‘s documentation and tutorials. The tutorials teach different uses of the toolkit, have been designed to be easy to follow, and come with the data necessary to follow them.
For more information or support on the use of TLK, contact us at firstname.lastname@example.org.