Microsoft researchers this week reported significant progress in their speech recognition efforts.
A year ago, the company attained a 5.9 percent word error rate in Switchboard, a conversational speech recognition task. Switchboard is a collection of recorded phone conversations that is commonly used by researchers to benchmark their speech recognition technologies.
On Aug. 20, Microsoft technical fellow Xuedong Huang announced his team had hit a human parity word error rate of 5.1, a new record.
As is increasingly the case at the Redmond, Wash., software giant, new advances in artificial intelligence (AI) has helped the company push the boundaries of what its software systems can accomplish.
“We reduced our error rate by about 12 percent compared to last year’s accuracy level, using a series of improvements to our neural net-based acoustic and language models,” stated the Microsoft researcher in an Aug. 20 announcement. “We introduced an additional CNN-BLSTM (convolutional neural network combined with bidirectional long-short-term memory) model for improved acoustic modeling.”
Additionally, Microsoft enlisted its suite of AI technologies called the Cognitive Toolkit (CNTK) in its record-setting bid. The company’s researchers also benefited from the performance boost provided by the GPU-assisted Azure virtual machines that use graphical processing units (GPUs) from Nvidia, similar to the video cards used by computer enthusiasts to generate high-fidelity graphics, to speed up AI workloads.
Although the 5.1 percent word error rate in Switchboard is a major accomplishment, the industry’s work is far from done.
Attaining humanlike levels of speech recognition in noisy environments using microphones at a distance remains a challenge. Accents and languages for which there is limited training data can also trip up voice-enabled systems.
Of course, Microsoft isn’t the only technology company exploring the growing market for speech-based computing.
Earlier this year, Nuance showed off an in-car speech recognition system that can help drivers and their passengers control an infotainment system without having to tap on a touch screen. Powered by AI, the company’s Dragon Drive software can intelligently control various components, like turning on a seat warmer when riders mention that they’re cold.
In April, Google announced the general availability of its Automatic Speech Recognition (ASR) service called Cloud Speech API (application programming interface). Already powering the voice capabilities in Google Search, Google Now and Google Assistant, the technology now boasts improved accuracy in long-form audio sessions and faster processing overall. Google also extended its file format support to include Speex, Opus and WAV files.
According to Google Product Manager Dan Aharon, businesses are already finding some novel use cases for the technology.
“Among early adopters of Cloud Speech API, we have seen two main use cases emerge: speech as a control method for applications and devices like voice search, voice commands and Interactive Voice Response (IVR); and also in speech analytics,” stated Aharon in an April 18 blog post. “Speech analytics opens up a hugely interesting set of capabilities around difficult problems e.g., real-time insights from call centers.”