Voice Recognition Technology

A touch of the future in the

Hon. Steve Rodan, in Douglas

Speaker of the the Hon Steve Rodan SHK explains how voice recognition technology is being deployed in the production of the Official Report in the Isle of Man.

Mr Rodan is the Speaker of the Clicking on the audio button House of Keys, the Manx revealed they had said: “they lower House. A Pharmacist, he pick up a £40 penalty”! was first elected to the Keys in 1995 and elected Speaker in The software 2, originally 2006. developed in Australia, relies on the use of individual voice WHEN it was suggested that profiles and these are speech recognition technology “harvested” – to use a technical might be able to produce expression – by asking each “instant” text for the Official Member to record a five- Report of 1, the idea minute pre-prepared passage. seemed as far-fetched as The resulting “profile” is then something out of a science improved over a period of time fiction novel. However, after by using the corrected audio two and a half years of hard from each session to adapt its work, science fiction became The downside acoustic and language elements reality in April 2008, when, There is, however, a caveat. – in other words, attempting to thanks to the vision and Whilst the use of our speech learn the Member’s speech perseverance of senior Hansard recognition software has so far patterns and use of language. editor Ian Faulds, the Isle of delivered up to 95% accuracy We anticipate that this process Man became the first for automatic transcription of will take approximately six Parliament in the well spoken English, it has months, but at the end of that Commonwealth to enter the done rather less well for heavy period of “training”, we expect instant transcription age. accents, slurred speech, sore the voice profiles to produce a throats or lisps. Remarkably, it much more accurate However, as you might expect, has coped well with Manx transcription. Of course, for it is a little more complex than Gaelic when the relevant words parliaments which sit more that! In simple terms Members’ have been given English frequently, the “training” of speech is digitally recorded and phonetics. The fact that the Members’ voice profiles could ‘converted’, through individual resulting automatic be considerably faster… voice profiles, into transcription is not 100% continuously scrolling text. accurate means that all Context is an important This is immediately available speeches do need to be edited element in speech recognition, to the Hansard editors at a before being published! Even a as the computer does not work remote location, so they can small error, such as omitting like the human ear. As part of begin the process of ‘tidying the word “not”, could be of the development process, we up’ what Members have said major importance when have created a substantial within three or four minutes of reviewing a debate! dictionary of Manx the beginning of a parliamentary phrases and parliamentary session. This has led to some light- expressions. In addition, as a hearted moments, particularly matter of routine, we now input The immediate benefit of the during the early stages of material from Order and system has been twofold: to development, when it quickly Question Papers prior to increase the speed of Hansard became apparent that editors debates taking place. All of this delivery to Members and the needed to remain alert to what helps to create the framework wider public, at the same time was really being said. Did, for of words within which the reducing the cost of example, the Member really speech recognition engine can production. For those of us say “a pig of a floating pouch successfully navigate. interested in connecting penalty”? Not quite. parliament to people these are important considerations.

The upside however, there has been a References: Although single user speech gradual but significant decrease recognition has been available in the number of self-employed 1. Tynwald, the Isle of Man’s for many years, what is new home-workers who have, parliament, consists of 35 about the Tynwald system is traditionally, provided the Members, 24 of whom sit in that it has successfully adapted backbone of the parliamentary the House of Keys and 11 in the technology to include an transcription service. From a the Legislative Council. Both infinite number of speakers on peak of 20 home transcribers in are Chambers in their own one audio channel. Each 2007, we expect to retain no right, but sit together as speaker has a separate more than 5 after completion of Tynwald for three days every microphone, and as he or she the transition to speech month. For further details of stands up to speak, the recognition in all three Tynwald and its Hansard look chamber editor uses a touch Chambers. at www.tynwald.org.im screen to switch in the correct voice profile, creating an The effect of these changes has 2. The speech recognition “utterance” which is then already been felt, with software, MultiSpeak, has been automatically transcribed. transcription costs for the short developed by Voice Perfect summer sitting of Tynwald (Australia) using a Dragon Meanwhile, back in their alone falling by more than speech engine, working offices, the other four full-time £7,000. Once speech together with VoicePower editors are already hard at recognition has been (UK) and Tynwald’s Office of work, correcting the text as it introduced into the House of the Official Report. appears on their screens. Keys3 and the Legislative Working consecutively, each Council4, annual savings are 3. The elected branch sits once editor takes an “utterance” and expected to exceed the total a week during the checks it against the audio, cost of the specialist software parliamentary term to consider which can be played and and associated hardware. legislation. replayed until he or she is satisfied with the result. From Although working with the 4. The revising Chamber sits taking an average of five or six new software has made once a week during the days to process and publish, we considerable demands on Mr parliamentary term. published our first same day Faulds and his editing team, Hansard on 15 April 2008 and who now work voluntary late now routinely achieve nights and sometimes sit at electronic publication within their screens almost non-stop one or two days. for up to a week, the end result has been well worthwhile – Clearly, this has been a particularly as the team feels it friendly revolution for the has touched the future! After Office of the Official Report, all, the concept of producing all of whose staff are fully Hansard throughout the committed to improving their Commonwealth can never be service to Members and the quite the same again. public. On a broader front,