17 Replies Latest reply on Mar 16, 2009 2:28 PM by joshtownsend

    Speech to Text: Is it working for you?

      Is anyone finding the speech-to-text function accurate enough to be useful? I was really looking forward to this feature, but have found that it transcribes speech so badly that it takes nearly as much time to correct it as it would to do a complete transcription. The audio quality of my source is very good -- a well-recorded studio interview. Is anyone having better luck? If you are, do you have any tips? I'm using a Mac, and transcribing American English.

      Thanks.

      David
        • 1. Re: Speech to Text: Is it working for you?
          akribie Level 2
          I ran a test on a very clear English voice with no background distractions and got a whole lot of words, only some of which related to what was actually said. Took a long time, too (around 1/3 real time from memory, but it was a few days ago so I might not be very accurate on that).

          Unless it does better in other languages, at this stage it looks more like a marketing tool than a serious feature.

          Win XP SP3, Core 2 Duo 2.8GHz, 2Gb RAM, PPro CS4
          • 2. Re: Speech to Text: Is it working for you?
            Jim_Simon Level 8
            I saw a review of the feature that suggested you might get a 50% success rate for the translation - with a good file, if you were lucky, the planets aligned, you found a four leaf clover on the way to work that morning...
            • 3. Re: Speech to Text: Is it working for you?
              Level 1
              So it looks like an interesting idea that is far from ready for prime time. If anyone from Adobe is watching this forum, Do you have anything to say about this "feature?" Is there any way to improve its performance?
              • 4. Re: Speech to Text: Is it working for you?
                Level 1
                This feature works best with "talking heads". Someone speaking into a good quality mic, in a clear measured pace. The best results I've seen were from audio that was recorded while someone was either reading or speaking from a prompter.

                That being said, I just read your post into my Macbook's built-in mic and got the following:
                "is anyone finding the speech to text function accurate enough to be useful I was really looking forward to this feature but it fell to the transcript speech so badly that it takes nearly as much time to correct it as it were to do a complete transcription the audio quality of my sources very good they will recording studio interview is anyone having better luck if you are do you have any tips I'm using a Mac and Trent spreading American English thanks David"

                Your mileage may vary... I selected "High Quality" for the transcription.
                • 5. Re: Speech to Text: Is it working for you?
                  akribie Level 2
                  My test with a very clean UK English talking head and High Quality was much closer to Jim's estimate of 50% than Chad's test. Maybe it is tuned for American accents?
                  • 6. Re: Speech to Text: Is it working for you?
                    TradeWind Level 3
                    There are options for different locales so you can choose the accent that most closely matches your source (not sure what happens when someone from the UK comes over to be interviewed on screen by an american...).

                    Adobe's recent comments seem to indicate that this feature was being developed for a much more thorough implementation in CS5, but when time came to release CS4, they figured the feature was at least useful enough in its current development that they included it anyway. They're very upfront and honest about it, saying early on that it isn't perfect, and will not always be useful for every kind of content.

                    I would expect finer controls and cleaner results in future revisions of the speech-to-text function.
                    • 7. Re: Speech to Text: Is it working for you?
                      Jim_Simon Level 8
                      >Maybe it is tuned for American accents?

                      American's don't have accents. It's everyone else that talks funny. unhappy smileys
                      • 8. Re: Speech to Text: Is it working for you?
                        akribie Level 2
                        Yes, I did set UK locale.

                        It is an interesting aside that has some limited use at the moment, but it's a bit cheeky to bill it as a major feature.
                        • 9. Re: Speech to Text: Is it working for you?
                          Level 1
                          It really is an interesting idea, but frankly, it works badly enough that it really ought to be billed as "experimental" or "demo" or something. In the case of my test, I was working with very clean recording made with a Schoepps mic in a quiet room; I used the "high quality, longer time" setting, and from what came out, I couldn't have even guessed the subject of the interview.

                          The fact that Chad Baker got as accurate a transcription as he did by reading into a mac mic makes me wonder if the system is somehow optimized for lower-quality audio sources. Is this possible? and would that mean some sort of pre-processing (compression, dynamic range, pitch or bandwidth limiting) would improve accuracy?
                          • 10. Re: Speech to Text: Is it working for you?
                            Curt Wrigley Level 4
                            Clear speech, clean noiseless recording, and constant pace seem to contribute to more accurate transcriptions. I have a file that is 90% accurate. I have others that are as waste of time. Its not a mature technology at this point. Useful only in very specific circumstances.
                            • 11. Re: Speech to Text: Is it working for you?
                              Level 1
                              Unfortunately, when it's as hit-or-miss as that, it can't be counted on and gets factored out of the workflow. But it's interesting enough technology that I'd be interested in beta (or is this alpha) testing. It is a truly great idea if it can be counted on.
                              • 12. Re: Speech to Text: Is it working for you?
                                Harm Millaard Level 7
                                Anyone who has worked with speech recognition software knows that you need to 'train' the software to recognize your voice, pronunciation, etc. to achieve anything over 90% accuracy. It is utter Utopia to expect PP to be able to recognize speech from interviews or anything else and turn that into usable text, without having been trained to recognize speech patterns, hesitations, pronunciations, voices, and other voice characteristics.

                                I predict that even in CS9 it will be of severely limited use. Imagine an interview in the US with a Frenchman, a Britain, an Aussie and a German being interviewed by an American. No way IMO that Adobe can make it work.
                                • 13. Re: Speech to Text: Is it working for you?
                                  watkino Level 1
                                  Well guys, I don't know what you expected, but for me, this feature is a Godsend and works MUCH better than expected, with at least 70%+ accuracy. The fact that I can look at the transcript, even with the occasional error, and get a good gist of the dialogue - WHAT A TIME SAVER!!! The time to process seems quite reasonable given what is being asked of the software.
                                  • 14. Re: Speech to Text: Is it working for you?
                                    Curt Wrigley Level 4
                                    When it works it is awesome. Searching for a word rather than scrubbing an hour long interview can save you bags of time. Temporal metadata is cool
                                    • 15. Re: Speech to Text: Is it working for you?
                                      ECS_1 Level 1
                                      I always do transcripts for interview features, the benefit they bring makes it crazy not to do them IMO. Even if Adobe can only give 60% accuracy it is useful if I can let it run overnight, as I could probably go through it and do corrections in realtime rather than the 1/5 of realtime which is what it takes me to do it manually.
                                      • 16. Re: Speech to Text: Is it working for you?
                                        I used this new feature on high quality, Pod cast studio with professional equipment and accuracy was not even close to 50%. Actually much less. Adobe must let consumer know, it is not ready for usage in actual commercial environment but perhaps for academic or experimental purposes only. A bit let down and could be deceiving if used in marketing campaign for this product.
                                        • 17. Re: Speech to Text: Is it working for you?
                                          joshtownsend Level 2
                                          "Anyone who has worked with speech recognition software knows that you need to 'train' the software to recognize your voice, pronunciation, etc. to achieve anything over 90% accuracy. It is utter Utopia to expect PP to be able to recognize speech from interviews or anything else and turn that into usable text, without having been trained to recognize speech patterns, hesitations, pronunciations, voices, and other voice characteristics."

                                          Totally agree with that. I think it's just a 'cool' feature to sucker people into getting CS4. At some of you find some use for it. For me it's mainly useless because I do stuff with actors and they tend to have emotion and slang in there voice which even with the best recording is useless.

                                          Unless of course they are really bad actors talk robots or people giving speeches.