Comment by rococode | Hacker News Reader

Comment by rococode | Hacker News Reader

Parent Full thread

I built a project where we transcribed speech into chat bubbles for an AR app (Magic Leap). We started with Google Speech-to-Text and swapped to Amazon Transcribe.

Performance-wise they seemed mostly similar - both can pick up text fairly accurately in somewhat noisy environments. We did notice that Google's offering seemed to be slightly more accurate. But the difference was marginal and only noticeable when we compared side-by-side on intentionally poorly pronounced words. There were two big differences that made it impossible for us to use Google's product:

1. Google has a 1 minute limit on streaming speech-to-text. It just closes the connection at 1 minute. It doesn't even send you a "final result", so if speech was being recorded at the time the connection dropped, that transcription is lost. Speaking of which...

2. Google doesn't provide incremental updates. So if someone speaks for a while, you only get an update at the end of it.

Note that this is their API - my impression is the product they use in their apps is superior in functionality to the product available on Google Cloud.

Amazon Transcribe, on the other hand, has a 4 hour limit and sends incremental updates, so in a longer sentence, like this one I'm currently writing, you would get a message every couple words, which is essential when the goal is to show a live transcription.

FYI google recently increased the limit to 5 mins. I found out when my streams stopped failing at 60 seconds :)

Also, they definitely have incremental updates.

>2. Google doesn't provide incremental updates.

Can you explain this comment? From my recollection, this statement is not true. Also, their site says:

"Returns text transcription in real time for short-form or long-form audio Cloud Speech-to-Text can stream text results, immediately returning text as it’s recognized from streaming audio or as the user is speaking. Alternatively, Cloud Speech-to-Text can return recognized text from audio stored in a file. It’s capable of analyzing short-form and long-form audio."

-https://cloud.google.com/speech-to-text/

Perhaps something has changed since I used it - about 2 months ago - or we were using it incorrectly, but in our experience it was "real-time" transcription, but not truly incremental. (Since you and another person have mentioned it's possible I'm leaning towards us not looking through the documentation carefully enough. We didn't spend too much time trying to figure it out because the time limit was a bigger dealbreaker)

Probably best explained as an example. Consider someone saying this:

"How are you doing? I'm doing well. Do you plan to go to the park with Alice and Bob tomorrow to see the fireworks show?"

When we were using it, Google would give:

1. "How are you doing?"

2. "I'm doing well."

3. "Do you plan to go to the park with Alice and Bob tomorrow to see the fireworks show?"

as three separate messages.

So it's certainly not waiting until the end of the entire streaming to give you a result - it sends you those three messages as you speak. In that sense it is "returning text as it's recognized", because once it recognizes you've finished a sentence it computes the words and gives them back to you. But the issue for us was that we could only get results after full sentences or long pauses.

Amazon, on the other hand, would give for the last sentence something like:

1. "Do you plan"

2. "Do you plan to go"

3. "Do you plan to go to the park"

So for our purposes (showing a chat bubble above a person's ahead), the latter version was much more useful. The reason is that in spoken language sentences often ramble, so we wanted to be able to show some incremental updates to the user as a long sentence was spoken so they wouldn't have to read 30 words at once.

From my memory (which is over a year ago), Google's service would give real-time results, likely meaning incremental word by word results. Did you use the gRPC interface or JSON? I think the streaming service is only available via gRPC.

Disclaimer: My comment reflects my own views, and not those of my employer, etc.

Live transcription limits are now 5 minutes.

Reply on news.ycombinator.com