Intelligence you can count on

FPinfomart monitors 50 Canadian radio stations in near-realtime.  This means you can browse and search for radio content broadcast 24 hours per day, 7 days per week, within about 30 minutes after it’s been aired.  The ability to access the complete contents of these channels, search the content by keyword, and instantly access audio clips through a self-service interface makes our service valuable.

The text attached to audio clips (transcripts) is generated using speech-to-text technology.  It is this computer-generated text which allows us to offer the complete contents of the monitored channels, as well as get the content online so quickly – there is no waiting for a human transcription service, or need to monitor only “high-priority” programs.  However, because the transcripts are generated automatically by computer, there can occasionally be unintended transcription “errors.”

The accuracy of the speech-to-text engine is actually quite good.  Mitigating factors which decrease the accuracy of transcripts include:

• Poor audio quality (e.g. an interview conducted over a static-prone phone connection)
• Heavily-accented speaker
• Use of words not in a dictionary (unusual names, less-common company names, slang, jargon, acronyms)
• Speaker using poor diction, or talking too quickly (making words run together)

Knowing this, and keeping in mind that there is no copy-editing of the automated transcripts, it is necessary to use different strategies when monitoring for radio content than you might use when monitoring the more strictly-controlled text in print and other “traditional media” sources.

You’ll want to use variations on spellings of the words you’re searching.  You’ll also want to think carefully about other English words that your keywords or phrases might sound like to a computer – especially if you’re searching names of people, places, corporations, or other entities.

The following video demonstrates two examples of the types of strategies you can employ to mitigate speech-to-text mistranslations in radio monitoring.