When machines speak, they sound stilted, robotic and mechanical 鈥� but they鈥檙e getting better. Google鈥檚 latest text-to-speech system, called Tacotron 2, generates sounds entirely from scratch, and the search giant claims the results are as good as those built using professional voice artists.
Previous systems normally produce speech by assembling human-recorded vocal sounds into words and sentences. In comparison, Tacotron 2 was trained on over 24 hours of human speech and corresponding transcripts, and could then generate completely new audio of phrases from a given text even if it had never seen some of the words before. You can .
at the University of East Anglia in the UK says the Google system is impressive because it learns all aspects of speech 鈥� including punctuation, prosody (the 鈥渢une鈥� of the voice) and intonation 鈥� without expert intervention.
Advertisement
Complex words
However, it is only the best examples that are really good, says 聽at the University of Helsinki. 鈥淚n longer utterances, the lack of understanding of the content of the text would likely be heard,鈥� he says.
The Google researchers admit that the system stumbles over complex words, such as 鈥渄ecorum鈥� and 鈥渕erlot鈥�, and can sometimes produce random strange noises. Plus, it can鈥檛 generate audio in real time, or be controlled to sound a certain way, such as happy or sad.
Other rival systems exist, such as Deep Voice produced by the Chinese internet company Baidu. The latest iteration, called is set to be unveiled at the International Conference on Learning Representations in April. 鈥淭he indications are that we are witnessing the birth of a new generation of speech synthesisers whose speech quality is virtually indistinguishable from human speech,鈥� says Cox.
While such systems are already improving voice assistants, they also raise concerns about falsified speech. Researchers at the University of Washington have manipulated video of former US president Barack Obama so that it appeared he was speaking words he never actually said, while startups such as Canada鈥檚 Lyrebird claim they can train their system to sound like anyone. Systems like these could be used for 鈥渟tealing someone鈥檚 voice鈥�, says Suni.
Reference: arXiv,
