Microsoft's conversational speech recognition system reached a record 5.1% error rate



Microsoft Artificial Intelligence and Research Lab based test which is at par with professional human transcribers with ability to listen to text several times, had earlier scored 5.9% error rate, and now, the conversational speech recognition system reached a record 5.1% error rate, which is perhaps the lowest so far.

The study transcribed recordings from the Switchboard corpus, that have been used by researchers to test speech recognition systems since the early 1990s, with a total collection of about 2,400 telephone conversations.

While the researchers at Microsoft AI and Research with also the same goal, achieved the same level of accuracy as the human transcribers who listened to what they were transcribing several times, access its conversational context and work with other transcribers.

The speech recognition system is used in services like Microsoft Cognitive Services, Cortana (the AI assistant), and Presentation Translator.

The Microsoft speech recognizer uses entire conversations, which let it adapt its transcriptions to context and predict what words or phrases were likely to come next, the way humans do when talking to each other.
Previous
Next Post »