AWS's Transcription Platform Is Now Powered By Generative AI (theverge.com) 13
Emilia David reports via The Verge: AWS added new languages to its Amazon Transcribe product, offering generative AI-based transcription for 100 languages and a slew of new AI capabilities for customers. Announced during the AWS re: Invent event, Amazon Transcribe can now recognize more spoken languages and spin up a call transcription. AWS customers use Transcribe to add speech-to-text capabilities to their apps on the AWS Cloud.
The company said in a blog post that Transcribe trained on "millions of hours of unlabeled audio data from over 100 languages" and uses self-supervised algorithms to learn patterns of human speech in different languages and accents. AWS said it ensured that some languages were not overrepresented in the training data to ensure that lesser-used languages could be as accurate as more frequently spoken ones. In late 2022, Amazon Transcribe supported 79 languages.
Amazon Transcribe has 20 to 50 percent accuracy across many languages, according to AWS. It also offers automatic punctuation, custom vocabulary, automatic language identification, and custom vocabulary filters. It can recognize speech in audio and video formats and noisy environments. With better language recognition, AWS said advances with Amazon Transcribe also bleed into better accuracy with its Call Analytics platform, which its contact center customers often use. Amazon Transcribe Call Analytics, now also powered by generative AI models, summarizes interactions between an agent and a customer. AWS said this cuts down on after-call work creating reports, and managers can quickly read information without needing to go through the entire transcript.
The company said in a blog post that Transcribe trained on "millions of hours of unlabeled audio data from over 100 languages" and uses self-supervised algorithms to learn patterns of human speech in different languages and accents. AWS said it ensured that some languages were not overrepresented in the training data to ensure that lesser-used languages could be as accurate as more frequently spoken ones. In late 2022, Amazon Transcribe supported 79 languages.
Amazon Transcribe has 20 to 50 percent accuracy across many languages, according to AWS. It also offers automatic punctuation, custom vocabulary, automatic language identification, and custom vocabulary filters. It can recognize speech in audio and video formats and noisy environments. With better language recognition, AWS said advances with Amazon Transcribe also bleed into better accuracy with its Call Analytics platform, which its contact center customers often use. Amazon Transcribe Call Analytics, now also powered by generative AI models, summarizes interactions between an agent and a customer. AWS said this cuts down on after-call work creating reports, and managers can quickly read information without needing to go through the entire transcript.
20 to 50 percent accuracy? Hmm.... (Score:2)
20 to 50 percent accuracy
At least that's better than the "English" manual that came with a recent electronics purchase....
Oops, TFA has been updated 20-50% improvement (Score:1)
The original article now says "Amazon Transcribe improves accuracy by 20 to 50 percent over its previous version across many languages, according to AWS."
An earlier version [archive.org] said "Amazon Transcribe has 20 percent to 50 percent accuracy across many languages, according to AWS."
Re: (Score:3)
Poor job of summarization.
Re:20 to 50 percent accuracy? Hmm.... (Score:4, Funny)
Poor job of summarization.
I think we found a race-condition bug :)
Re: (Score:2)
Re: (Score:2)
Your request for a free pentest of your site has been granted. Results will be posted here for your review and everyone's enjoyment.
Is this really new? (Score:2)
Hasn't voice recognition been using trained neural network models since the beginning? This would just be another form of that.
Speech... (Score:2)
WOw (Score:2)
"Amazon Transcribe has 20 to 50 percent accuracy across many languages" - that's not a transcription then, it is trully generative AI's interpretation of the content. I'm sure all who use translation services will flock to this garbage :)
I'm not impressed (Score:2)
Amazon Transcribe has 20 to 50 percent accuracy across many languages
I can already have that with any manual for any cheap Chinese crap I buy. In other words, that junk is no better than an underpaid and overworked Chinese worker with a half-assed second language education.
Audio recorded from Alexa devices? (Score:4, Interesting)
Gee, I wonder where they got their massive dataset of unlabeled speech... ?