Microsoft Officially Makes First Humanly Accurate Speech Recognition Tech
Microsoft will herald a global change in the realm of speech recognition. Its newest software appears to transcribe conversational speech as good as humans.
This is a historic achievement for the company, as a study just revealed that it can best professional human transcriptionists in conversational speech. It has a word error rate of 5.9 percent, lower than 6.3 percent just last month. Microsoft said this is the lowest ever recorded in the industry-standard Switchboard task, and it's even lower than the errors of human professional transcriptionists.
Xuedong Huang of Microsoft said this means Microsoft has achieved "human parity." Their new technology uses neural language models which helps efficient generalization by grouping similar words.
According to Futurism, the achievement came decades after speech pattern recognition was first studied in the 1970s. Given that Google's DeepMind is not far behind in the same field, Microsoft's achievement will herald more developments in the field of artificial intelligence research.
This also means the new technology can be used for Microsoft's personal voice assistant Cortana for Windows and Xbox One. Aside from better speech-to-text transcription, it can help make Cortana more powerful and a "true" intelligent assistant.
It can be remembered that other voice assistants exist, starting with the more popular Siri from Apple. Recent achievements from DeepMind include a "voice synthesis program" that can generate its own human-like speech.
Still, Microsoft clarified that this didn't mean perfection. Their technology still cannot recognize words clearly, which something even humans cannot do efficiently. Regardless, their next goal is to make computers "understand" human conversation. Ergo, a transfer to the realm of understanding from recognition.
However, this possibility appears to be very likely considering how the realm of artificial intelligence is already making leaps and bounds in various fields. For instance, it's reported that Google's DeepMind already has about 1,000 ongoing AI projects.