Software turns English into synthesized Chinese almost instantly.   November 8, 2012

Why It Matters

Software that instantly translates spoken speech could make communicating across language barriers far easier. We’re rapidly moving into a world of closer relationships with technology offering an exciting possibility for self expression and transformation.

It could be the next best thing to learning a new language. Microsoft researchers have demonstrated software that translates spoken English into spoken Chinese almost instantly, while preserving the unique cadence of the speaker’s voice—a trick that could make conversation more effective and personal.

The first public demonstration was made by Rick Rashid, Microsoft’s chief research officer, on October 25 at an event in Tianjin, China. “I’m speaking in English and you’ll hear my words in Chinese in my own voice,” Rashid told the audience. The system works by recognizing a person’s words, quickly converting the text into properly ordered Chinese sentences, and then handing those over to speech synthesis software that has been trained to replicate the speaker’s voice.

Video recorded by audience members has been circulating on Chinese social media sites since the demonstration. Rashid presented the demonstration to an English-speaking audience in a blog post today that includes thisvideo.

Microsoft first demonstrated technology that modifies synthesized speech to match a person’s voice earlier this year (see “Software Translates Your Voice Into Another Language”). But this system was only able to speak typed text. The software requires about an hour of training to be able to synthesize speech in a person’s voice, which it does by tweaking a stock text-to-speech model so it makes certain sounds in the same way the speaker does.

AT&T has previously shown a live translation system for Spanish and English (see “AT&T Wants to Put Your Voice in Charge of Apps”), and Google is known to have built its own experimental live translators. However, the prototypes developed by these companies do not have the ability to make synthesized speech match the sound of a person’s voice.

The Microsoft system is a demonstration of the company’s latest speech-recognition technology, which is based on learning software modeled on how networks of brain cells operate. In a blog post about the demonstration system, Rashid says that switching to that technology has allowed for the most significant jump in recognition accuracy in decades. “Rather than having one word in four or five incorrect, now the error rate is one word in seven or eight,” he wrote.

Microsoft is not alone in looking to neural networks to improve speech recognition. Google recently began using its own neural network-based technology in its voice recognition apps and services (see “Google Puts Its Virtual Brain Technology to Work”). Adopting this approach delivered between a 20 and a 25 percent improvement in word error rates, Google’s engineers say.

Rashid told MIT Technology Review by e-mail that he and the researchers at Microsoft Research Asia, in Beijing, have not yet used the system to have a conversation with anyone outside the company, but the public demonstration has provoked strong interest.

“What I’ve seen is some combination of excitement, astonishment, and optimism about the future that the technology could bring,” he says.

Rashid says the system is far from perfect, but notes that it is good enough to allow communication where none would otherwise be possible. Engineers working on the neural network-based approach at Microsoft and Google are optimistic they can wring much more power out of the technique, since it is only just being deployed.

“We don’t yet know the limits on accuracy of this technology—it is really too new,” says Rashid. “As we continue to ’train’ the system with more data, it appears to do better and better.”

 Other Options

This from my friend Allen B.: