SpeechBot

1999 - 2005
dead us
compaq's research lab put speechbot on the web at the end of 1999, running speech recognition across public radio archives so you could search the words inside an audio file. almost nothing else was attempting it. the transcripts were wrong often enough that the site said so on the front page, and they were still good enough to drop you at the minute of a show where somebody said the thing you half remembered. hp inherited the project along with compaq, kept it running a few more years and switched it off in 2005 when the cambridge lab closed.

~ the obit ~

i work on machine vision now, which mostly means building models that read something a person would otherwise have to sit and watch. speechbot was that idea pointed at audio and it shipped in december 1999. the uncomfortable part is not that it existed, it is how ordinary the description sounds today and how completely alone it was at the time.

compaq launched it at streaming media west in san jose, out of the corporate research lab in cambridge, massachusetts, with a group on the gold coast in queensland working the recognition side. the lab was the one compaq had inherited from digital, which is a lineage worth knowing when you look at what it was running on.

what it indexed was american public radio, and that is my honest distance from it. fresh air, car talk, the diane rehm show and the pentagon's daily briefings were not my listening. by the time i had a connection that could pull down hours of streamed talk, the site was in its last year and i had never heard of it. the thing i am writing about here is the machine, not a habit.

~ the rap sheet ~

Borndecember 1999, launched at streaming media west in san jose
Peakover 17,000 hours indexed by mid 2003
Killed2005, last serving on 25 october
Lifespanjust under six years
Made bycompaq corporate research, cambridge massachusetts, then hp labs
Killed bythe closing of the lab that owned it

~ how it worked ~

the pipeline was assembled from whatever the lab happened to own. windows workstations sat there pulling the streams down off the public archives. a linux cluster ran the recognition, using a recognizer called calista that descended from carnegie mellon's sphinx-3. alphaservers on tru64 unix held the web server, with the search index and the metadata library on the same machines. it was a research group's hardware closet turned into a production service.

the output was a rough transcript, and speechbot never pretended otherwise. the front page said in plain words that transcripts based on speech recognition are not exact. what made it useful anyway was the linking: a hit did not just tell you which show contained your words, it handed you a stream offset, so you clicked and the audio started a few seconds before the moment. searching inside a two hour broadcast was the entire product.

~ what was in the index ~

the last working front page, captured on 1 september 2005, says speechbot currently indexes 15321 hours of content, and lists the twenty shows it was pulling from under seven headings. fresh air under arts and entertainment. american radioworks, here and now, on point and the pbs newshour under current events. afrts radio news under government and military, alongside the white house feed and the department of defense briefings. the motley fool and marketplace under personal investment. car talk, the brian lehrer show, the charlie rose show, the connection, the diane rehm show and the leonard lopate show under talk. scuba radio under sports, which is the one that makes the list feel like a real thing somebody maintained by hand.

the number had been higher. hp's own writeup in mid 2003 put it over 17,000 hours, so the index shrank by roughly two thousand hours over the last two years of its life. shows moved their archives or changed to formats the crawler could not reach, and nobody was adding new sources fast enough to cover the loss. an index that gets smaller is usually a sign that the people running it have been assigned to something else.

~ why it stopped ~

hp acquired compaq in 2002 and speechbot came with the furniture, renamed on the page from compaq.com to hp speechbot and rehoused inside hp labs navigation while keeping its old compaq address to the end. then hp reorganized its research in 2004 and 2005, folding the individual centers into hp labs, and the cambridge lab that had built the thing was among the casualties. the service had no owner after that.

the archive gives the last day it worked. speechbot.research.compaq.com is captured normally through october 2005, at about 6,800 bytes, with the final one on 25 october. there is nothing after it. no farewell page and no redirect to a replacement. a search engine that had spent six years listening to american radio simply stopped answering, and the closest thing anybody wrote about it was a trade note observing that it was no longer available.

~ leave a rose ~

an anonymous tribute. one rose per visitor per day.

@}-->-- ... roses left here

~ leave a wreath ~

a rose fades by morning. a wreath stays on this grave for good, with your name on it.

paid once through stripe. payments are handled by OÜ Lakeview, the small company that keeps rip.so running. the money pays for the hosting and the nightly digging. adopt it and the headstone reads "maintained in memory by" you. the groundskeeper plants new wreaths at the next tending, usually within a day.

» one email a week: new graves and the anniversaries of the dead. subscribe

~ leave a note ~

visitors before you have left these graveside notes. anonymous welcome.

other search: Arianna, Daypop, Ms. Dewey, Qwant Music

also died in the 2000s: Deepnet Explorer, Hitel, Buena Vista Internet Group

« previous  ·  random  ·  next »