Search Engines Break the Sound Barrier

Computerworld - Do your telemarketers consistently make legally required disclaimers when selling securities? If your firm records its telemarketing calls, IT could set up audio mining software to let management search audio file archives to quickly find the answer.
Emerging audio mining tools, also called audio indexing or audio search software, offer speech processing and search technologies in a single package. The speech engine creates an index that includes a time and date stamp for each spoken word or phoneme in an audio or video file. The search engine then uses that index to allow rapid identification and playback of specific passages. The software may also apply metatags that identify the speakers or the subject of a given passage.
The speech-processing accuracy of speech-to-text engines, traditionally used to index high-quality broadcast audio, has advanced to the point where vendors are introducing new packages for indexing more informal conversations, ranging from corporate meetings to training videos and even help desk telephone conversations.
"[The technology] seems to have passed the threshold of usability," says William Meisel, president of TMA Associates, a speech-recognition consulting and market research firm in Tarzana, Calif.
Unlike speech-to-text packages, which can be trained for individual users, audio indexing products are speaker-independent. They also rely on large, language-specific vocabulary dictionaries, as well as domain models that may optimize for the type of conversation (e.g., telephone) or industry (e.g., health care). While the newest products can process audio at or faster than real time with an accuracy sufficient for searching, the output text isn't a readable transcript, cautions Jackie Fenn, an analyst at Stamford, Conn.-based Gartner Inc. And as new companies, products and terms come into use, users must update their systems regularly or face what Francis Kubala, division scientist at Cambridge, Mass.-based BBN Technologies, calls "the out-of-vocabulary problem."
Audio mining's most compelling fit may be for applications where a searchable index can replace the need for transcription. In contrast, data mining of audio content for marketing purposes is "a little bit of an evangelistic sell" at this point, Meisel says.
"The call center is a little tougher, because you may or may not discover something [with audio mining]," explains Fenn.
The technology's greatest value may be derived from embedding it in other applications. San Mateo, Calif.-based Virage Inc., for example, offers both Atlanta-based Fast-Talk Communications Inc.'s Fast-Talk and BBN's Audio Indexer as plug-ins to its VideoLogger video indexing system. More advanced applications could eventually integrate call center logs with sales activity and other customer relationship management data, analysts say.
But audiomining hasn't worked in every case. Ted Ryan, manager of collections development at Atlanta-based The Coca-Cola Co., says he wanted to use it to index television commercials last year, but "the voice-overs clashed with the music." With an accuracy rate of just 15%, he turned to manual transcriptions.
Coca-Cola also tried using audio indexing of meetings. "Our chief executive [at the time] was Cuban. When we ran it with executive speeches, it came up with gobbledygook," Ryan says.
Nonetheless, he says he's interested in testing the latest tools to index radio advertisements. And accuracy continues to improve, says Kubala, adding that he expects the word error rates for nonbroadcast audio to drop dramatically during the next three years.

Read more about App Development in Computerworld's App Development Topic Center.