4 ms·

I tried this a few weeks ago with Watson for that purpose (being able phase out during conference calls); complete gibberish came out on all my trials (about 40

I tried this a few weeks ago with Watson for that purpose (being able phase out during conference calls); complete gibberish came out on all my trials (about 40 conference calls) with 2,3,4,5 and 8 people, all in English. All come out complete garbage. You can't even guess what we are talking about reading the entire transcripts. Not sure if this is not just the same thing... If it is, I would love to hear your experiences. Edit: shame I didn't publish it :) Edit2: I had one call for which I have a colleague speaking who speaks 'the Queen's English' and same result...

adrianN 10y ago
You know what they say, garbage in, garbage out.
ldehaan 10y ago
wouldn't that just mean that Watson sucks at text to speech? also since that's probably the case for most tts services (at least in my experience) why not run the audio through all the tts engines you can find and compare results. keep what has the highest matches and see if you can get the other words using phonetics engines. for words that are specific to your company train a tts engine on those words and weight your local engine higher than other services. you can get a pretty decent translation back, but who would be so lazy as to work that hard..... ;)
- skykooler 10y ago
  A large part of the issue is that if you are dealing with phone meetings, audio is severely compressed, and not very similar to the training data that Watson has been taught with. Google Voice might be able to do a better transcription, because they have much more data from phone messages to work with, but eventually it comes down to the fact that voices over a phone line are a lot harder to understand without context processing than voices recorded with a decent microphone.
caseysoftware 10y ago
It's how they've structured their audio for ASR. When you have a single track for all channels, the signals get garbled and it's hard to segment each out. The key is using a conference call system which captures each track separately. Polycom, Tropo, and others do it. Then when you run ASR, you can feed in each channel and have a "unified" transcript. And yes, I've worked on this problem at my last startup where we did all of that and took things further by adding search. Here's a demo: http://clarify.io/try-it-now/?id=20f5e8c4f1ef4f8c91160839b4815eb1 http://clarify.io/try-it-now/?id=20f5e8c4f1ef4f8c91160839b48...

adrianN 10y ago

You know what they say, garbage in, garbage out.

ldehaan 10y ago

wouldn't that just mean that Watson sucks at text to speech? also since that's probably the case for most tts services (at least in my experience) why not run the audio through all the tts engines you can find and compare results. keep what has the highest matches and see if you can get the other words using phonetics engines. for words that are specific to your company train a tts engine on those words and weight your local engine higher than other services. you can get a pretty decent translation back, but who would be so lazy as to work that hard..... ;)

skykooler 10y ago
A large part of the issue is that if you are dealing with phone meetings, audio is severely compressed, and not very similar to the training data that Watson has been taught with. Google Voice might be able to do a better transcription, because they have much more data from phone messages to work with, but eventually it comes down to the fact that voices over a phone line are a lot harder to understand without context processing than voices recorded with a decent microphone.

caseysoftware 10y ago

It's how they've structured their audio for ASR. When you have a single track for all channels, the signals get garbled and it's hard to segment each out. The key is using a conference call system which captures each track separately. Polycom, Tropo, and others do it. Then when you run ASR, you can feed in each channel and have a "unified" transcript. And yes, I've worked on this problem at my last startup where we did all of that and took things further by adding search. Here's a demo: http://clarify.io/try-it-now/?id=20f5e8c4f1ef4f8c91160839b4815eb1 http://clarify.io/try-it-now/?id=20f5e8c4f1ef4f8c91160839b48...