Customer-centric companies depend on Speech Analytics to understand and predict customer behavior in ways never before possible."
- Company Profile
- Phonetic Search Technology
- Phonetics Advantages
- Intellectual Property
- News & Events
- In The News
- Upcoming Webinars
- All Archived Webinars
- Community and Awards
- Board & Management
- Executive Management
- Technical Management
- Board Members
Avid PhraseFind: 10 Tech Notes
By: Michael Kammes
Apr. 06, 2011
PhraseFind is a phonetic matching algorithm, developed by Nexidia. It’s been wildly successful, serving as the basis of AV3’s “Get” software, which is used in conjunction with Final Cut Pro. Avid has licensed the technology and has branded it “PhraseFind”. It is available for $495, or $1295 when bundled with ScriptSync. Additional language packs can be purchased for $149.
See Avid’s indepth FAQ here: http://avid.custkb.com/avid/app/selfservice/search.jsp?DocId=389887&ssdFilterCommunity15=1408&ssdFilterCommunity10=369&ssdFilter_SearchKeyWord=&Hilite=#a10
Since this is the first incarnation of Avid integration, there are a few gotchas.
1. PhraseFind is currently only supported with Avid Media Composer 5.5.1 (Mac or PC), Symphony 5.5.1 (Mac or PC) or Newscutter 9.5.1 (PC). Earlier versions are not supported.
2. The database Phrasefind creates cannot currently be shared. Thus, every station accessing files must independently index the files. Via Avid: In each Project folder, you will see a new folder called SearchData which contains a file called _SearchDB_. This file is the database created by Find when it indexes the bins within your project. If you have PhraseFind installed, you will also see a folder called PhoneticData which contains a number of files with a .pat extension; these files contain the phonetic information for each of the audio clips within your project. One .pat file will be created per clip. DTL – can you say workgroup version?
3. Linking via an AMA file or even importing does NOT cause PhaseFind to index the file. The Bin must be saved before PhraseFind begins.
4. PhraseFind indexing also does not appear to be multithreaded. Routinely, one processor can be seen (barely) being taxed during indexing.
5. In my tests, anything Phrasefind search returns over 60% probability seems to usually be spot on. This decreases sharply as the audio fidelity and annunciation decreases. DTL – BING! Needs a filter.
I used Barack Obama’s “A More Prefect Union” speech for accuracy testing. This offered an excellent testing opportunity, as I not only had the full speech, but also a word for word transcript of the speech to compare against. I searched for the word “Dreams”. The transcript from the speech told me the word “dreams” was used a total of 5 times. With PhraseFind, 15 hits were returned, all above 50%. The top 6 hits were ranked as 56% accurate or better. They all turned out to be correct, including a hit for the word “dream” (singular).
Next, I tried the word “education”. It was spoken 6 times, and PhraseFind found all 6 instances at 85% or higher degree of certainty.
6. Moving the files from one bin to another, as well as renaming the bin, causes PhraseFind to momentarily rewrite it’s database. It’s not re-indexing the files, it’s simply rewriting it’s database.
7. If a clip is deleted from the project, the associated PhraseFind entry is also deleted.
8. Approx 13minutes = 1MB for a PhraseFind .pat file.
9. Get used to NOT hitting “RETURN” on the keyboard after bringing up the SEARCH box. That searches just text. You need to click on the PhraseFind button. DTL – this is another difference between GET and PF, the integrated dialog and metadata search.
10. Index time is impacted by AMA linking vs native DNxHD (MXF) files. This seems to vary wildly depending on the codec associated with the AMA file.