MetService has replaced actors providing its phone-based weather reports with text-to-speech technology, having sent Maori staff to the UK to help the Scottish software cope with local place names.
The organisation’s MetPhone service was launched in 1992, using a system from Auckland-based Amtel Communications. It used actors and radio announcers to supply voice-overs for the interactive voice relays on the push-button service.
But this limited the number of forecasts that could be given, and after a while the 15 or so part-timers would “get tired and their voice quality would drop”, says operations manager Stephen Harris.
MetService realised text-to-speech technology would be inevitable and in 2000 began seeking systems. Some were tested but MetService believed the systems were not mature enough, so the project was put on hold for six months. In 2001 it came across Rhetorical from Scotland, one of whose voices sounded similar to one of MetService’s own voice loaders.
Harris says the rVoice software chosen needed some “tweaking” for New Zealand words as it was delivered in an English voice.
Phonetic spellings used Rhetorical’s phonetic dictionary, though a few of MetService’s Maori staff were sent to Scotland to make recordings of Maori place names.
MetService also sent over some 500 common words, 300 Maori and 200 European, for further “tweaking” as the system wasn’t pronouncing them correctly. Rhetorical sent back recommendations based on the phonetic spelling.
Harris says there is no perfect pronunciation for many places, such as Paraparaumu, as accents differ. The aim is to make the recordings understandable so the agency can get its message across. Making them too precise may make certain Maori place names incomprehensible to some Europeans.
MetService says its system searches for keywords, so it can cope with spelling mistakes and automatically correct them. This system can also detect rude or unsuitable words supplied to it from third parties.
The Rhetorical system went live on July 1 last year, but was taken off the following day as MetService felt there were “enough little glitches and the words did not sound right”.
But after a month of further “tweaking”, the system was relaunched, with MetService saying it has operated flawlessly since then. The project’s cost was below $100,000, Harris says.
When MetService tested the software, it sought assurances from other Rhetorical users in Australia and elsewhere — Harris is unaware of any in New Zealand — who spoke “glowingly” of the product. However, he warns that automated systems need constant monitoring.
“When you have a human you can see whether they are working, but how do you know a piece of software is working? We have two text-to-speech systems, one feeding an archive and one feeding live. They are constantly monitoring each other. We send 1200 files on a scheduled basis to MetPhone and if they do not arrive, alerts are sent to the forecasters. We still have forecasters at the front end.”
Harris believes such text-to-speech technology has added uses in AA-type road reports, snow forecasts and regional council warning systems, particularly where text has to be converted to voice quickly and at unusual times. He advises anyone looking at such a system to plan for instant correction as things can automatically go wrong as well as right.