In an earlier post, I covered the emergence of audio-search and audio deep tagging players like Pluggd, Podzinger and Blinkx. The latest issue of Wired provides a better understanding of how Pluggd is applying a Google-like indexing approach to audio content.
Acoustic speech recognition technology has been around for years — companies such as Podzinger, blinkx and Podscope have used it to build similar products — but it’s only part of the Pluggd picture. To understand Pluggd’s potential, you need a quick peek under the hood.
First, the company uses parallel servers to churn through audio, performing a speech-to-text analysis of each file at faster-than-real-time speeds. The company also maintains what it calls a "concept map," a database that tracks associations between words by analyzing Pluggd’s speech-to-text transcripts and looking for words that often show up in the same contexts. This tool also studies text web pages, so the concept map learns to associate words faster than Pluggd can analyze audio files.
Because it can identify related concepts — or a lack thereof — appearing near your search term, Pluggd uses the associations to home in more closely on promising results. The company told me this reduces its error rate to "close to zero." Considering the processing requirements of analyzing this much audio and text, it’s not surprising to learn that Intel was an early investor in Pluggd.