3 ms·

"We think we have created one of the best speech recognition engines!" Then why don't you have any numerical evidence for that claim? Speech processing/recogni

"We think we have created one of the best speech recognition engines!" Then why don't you have any numerical evidence for that claim? Speech processing/recognition isn't a new field by any means, so you should be able quantify performance for your particular domain. I don't know if any such numerical results would drive conversions for you, but it should help you get something more definitive than "we think".

PaulHoule 11y ago
Speech recognition is an area where objective evaluation has been as much part of the problem as the solution. If the paradigm is "get the highest recognition rate over sample X" you can have super-human accuracy but the product still sucks because you are saying "Xbox Play Titanfall" over and over again and it never figures it out because it doesn't engage you in a dialogue to understand what you are saying. We've seen the same problem in full text search. The TREC evaluations have, over the long term, taught us how to make better search engines, but if you looked at any one year of TREC you would get depressed about the prospect of making a better search engine because you'll see all the things people tried that failed. When Google came out the TREC community was overturned because it was clear you could make a much "better" search engine, but measured the way TREC measured things at the time, Google would not have appeared better, because TREC rewards you for being accurate at position 1000 and Google cares mostly about position 1. Customers perceive the latter, but they will never get to 1000 unless they are researching a patent application or something.
- fundamental 11y ago
  To an extent I have to agree. Trying to figure out how to measure success is painful and each objective is going to have some fault. The key here is "for your particular domain". The particular target of "voice journals" would indicate single individual, relatively low background noise, microphones which could be found in common smartphones, etc. Per general recognition, success is going to be biased towards successfully identifying words which would actually be searched for. Given the high price having some sort of evaluation beyond do it yourself would seem reasonable to me. I find it somewhat hard to take something seriously which describes itself with: - "Finally a high quality X" (relative to what?) - "We think" (that sure doesn't make it sound like you've done your research) - "trained on countless hours of audio" (Massive understatement given typical speech recog. projects) - "best possible accuracy" (ok, where/how have you measured that) I haven't gone into the full text search field, but I've seen plenty of areas where measuring success is a convoluted mess, which doesn't necessarily lead to better numbers in real world scenarios. Computer vision in particular seems to have plenty of these local optima scattered around.

PaulHoule 11y ago

Speech recognition is an area where objective evaluation has been as much part of the problem as the solution. If the paradigm is "get the highest recognition rate over sample X" you can have super-human accuracy but the product still sucks because you are saying "Xbox Play Titanfall" over and over again and it never figures it out because it doesn't engage you in a dialogue to understand what you are saying. We've seen the same problem in full text search. The TREC evaluations have, over the long term, taught us how to make better search engines, but if you looked at any one year of TREC you would get depressed about the prospect of making a better search engine because you'll see all the things people tried that failed. When Google came out the TREC community was overturned because it was clear you could make a much "better" search engine, but measured the way TREC measured things at the time, Google would not have appeared better, because TREC rewards you for being accurate at position 1000 and Google cares mostly about position 1. Customers perceive the latter, but they will never get to 1000 unless they are researching a patent application or something.

fundamental 11y ago
To an extent I have to agree. Trying to figure out how to measure success is painful and each objective is going to have some fault. The key here is "for your particular domain". The particular target of "voice journals" would indicate single individual, relatively low background noise, microphones which could be found in common smartphones, etc. Per general recognition, success is going to be biased towards successfully identifying words which would actually be searched for. Given the high price having some sort of evaluation beyond do it yourself would seem reasonable to me. I find it somewhat hard to take something seriously which describes itself with: - "Finally a high quality X" (relative to what?) - "We think" (that sure doesn't make it sound like you've done your research) - "trained on countless hours of audio" (Massive understatement given typical speech recog. projects) - "best possible accuracy" (ok, where/how have you measured that) I haven't gone into the full text search field, but I've seen plenty of areas where measuring success is a convoluted mess, which doesn't necessarily lead to better numbers in real world scenarios. Computer vision in particular seems to have plenty of these local optima scattered around.