By STEPHEN COX

The objective of the TC-STAR project was ambitious: to achieve a breakthrough in SST that significantly reduced the gap between human and machine translation (MT) performance.  Funded by the European Commission as part of its Sixth Framework Program, it recently finished after three years of research. 

TC-STAR was unusual in that it sought to develop and integrate three major areas of speech and language processing: speech recognition, machine translation and speech synthesis.  A consortium of 13 leading laboratories in these fields from Europe and the USA was assembled to address this challenge.  The architecture of the system used in TC-STAR is shown below:

The system was developed and tested on material from the European parliament in Spanish and English and from Voice of America in Chinese and English.

TC-STAR was also unusual in that it adopted a DARPA-style competitive strategy, with partners developing technology components and then comparing their performance in competitive evaluations, also open to external participants.  A workshop, organized after each evaluation, facilitated sharing to all participants of the techniques and ideas shown to be most successful.  This approach seems to have paid off: Figure 1 shows the word error rate for speech recognition in TC-STAR and Figure 2 the BLEU score for machine translation (BLEU is an automatic method for scoring the quality of translations that has been shown to correlate well with human judgements).  And end-to-end system was produced for English/Spanish and was evaluated by subjects in terms of its fluency, naturalness, listening effort and overall quality. 

Although there is still a long way to go in SST before we can produce systems that approach human quality, the project leader Gianni Lazzari of ITC is positive: "A key area that TC-STAR addressed was the interface between speech recognition and language translation: research in areas such as confusion network coupling, automatic segmentation and punctuation gave us the ability to build fully-fledged speech translation systems that performed reasonably well. TC-STAR has also revealed many research topics to explore in order to improve the translation quality: for instance, introducing linguistic knowledge in the statistical framework, new learning algorithms, new and more powerful automatic metrics."

Figure  1: Speech recognition word-error rates over duration of TC-STAR project

Figure  2: BLEU scores over duration of TC-STAR project