The Environment Agency and Natural Resources Wales have switched automated flood warning phone calls to Amazon’s Polly service, claiming both clearer output and cost savings.
The agencies have more than 2 million telephone numbers registered across England and Wales for automated calls warning about flood risks. The system converts text entered by local duty officers into audio, with more than 2,000 hours generated in 2018.
However, the results of the old system sounds “unnaturally electronic with often poor pronunciation and cadence,” writes Ben Smedley, external digital services incident management and resilience at the Environment Agency, in a blogpost. It also requires a large custom dictionary, mainly of place names, and the payment of an ongoing licence free for the text-to-speech software.
Smedley says they looked for alternatives that could use machine learning and neural networks to provide a better service, including custom pronunciation for place names and a Welsh language and accented service that works near instantly. After extensive testing with a range of suppliers, the agencies selected Amazon’s Polly service, which has an English voice called ‘Amy’ and a Welsh one, ‘Gwyneth’.
To demonstrate the difference, the agencies published the same message generated by the old system and the new one. While neither sounds entirely natural, the new one is clearer on specific words such as ‘Richmond’ and ‘year’ and has fewer examples of odd intonation.
The new system is also much cheaper, with costs from the text-to-speech translation service falling from more than £40,000 to less than £1,000 annually.
“We now have a far more flexible and powerful service that will only get better as the Amazon Polly neural network technology evolves and improves,” writes Smedley.
In August, the Environment Agency started supplying its flood alerts to Google, so these could appear on its mapping services.