AssemblyAI
The best way to build Voice AI apps with one robust API
Start new thread

Universal-1 - Multilingual speech AI model trained on 12.5M hours of data

by•
Try AssemblyAI's most capable and highly trained speech recognition model trained on 12.5M hours of multilingual audio data. Universal-1 achieves best-in-class speech-to-text accuracy, reduces word error rate and hallucinations, and improves timestamps.

Add a comment

Replies

Best
Maker
📌
The Universal-1 Speech-to-Text model was created to focus on the nuances of spoken language across accents, tone, dialect, faithfulness, and more. We hope the new capabilities of Universal-1 will help power the next generation of AI products and features built with voice data. Give it a try and let us know what you're building!
Awesome! How close is it to human level interpretation? Are we 75% there or? What's the benchmark? Or is there a benchmark?
Maker
Great question! Universal-1 in English came in at 92.7% accuracy across multiple datasets. We display all benchmarks here: If you'd like to see a more in depth analysis of Universal-1, I highly recommend you check out our research which includes Word Error Rate by language, timestamp accuracy, hallucinations, and more. > Hope this helps! :)
I think there could be a clearer benchmark 92.7% against what?. I.e: Humans can interpret gibberish, or very chopped up or noisy audio. By filling in the blanks, or using context.
Cool!👍 The free version can transcribe up to 100 hours of audio, It's sufficient for the initial use of the project.
Congratulations on the launch! How did you manage to ensure high accuracy for languages with fewer training data?
Joy to have this app!!!
Congratulations on the launch 🎊 I think this can be used for converting podcasts into blogs.🤔
congratulations on the launch of universal-1, dylan, britney, and meredith. your focus on reducing word error rates and enhancing dialect recognition is impressive. could you share how universal-1 handles low-resource languages and if there are plans to expand its linguistic dataset?
Seems like a great product for all the devs. Congrats on the launch!
Excited to test this. Does the playground currently allow us to access this model? Or will it soon? I'm assuming part of what makes this work well is the same process you've been using, with word prediction based on context, as opposed to straight word for word phonetic reproduction of what it "hears"?
Maker
Yes, the playground is currently utilizing Universal-1! You will be able to select tiers, and Universal-1 currently falls under [Best]. More on our tiers and research here:
appreciate the reply and confirmation. I assume this probably works well for cleaner audio, but I'm finding newer transcription iterations, including universal 1after testing, are actually regressing for muddier audio. Their tendency to rewrite and predict what was supposedly said, as opposed to just trying its best to phonetically reproduce word for word, is resulting in compelling looking but pretty far off results. Is there a way to adjust for this on our end?
How good is it at Korean, cause elevenlabs is bad at that.
Maker
Our Best tier supports Korean and includes speaker labels, custom spelling, automatic punctuation, etc. We encourage you to try our API for free - more on supported languages and account creation here:
Loved it. seems to be one of a kind.
12
Next
Last