The "universal translator" which was part of Star Trek and a lot of other Sci-Fi I was exposed to as a kid was something I was really fascinated with. My Dad worked as a simultaneous French->English translator and sadly spent long hours away from home and, as a kid, I started trying to build a translator so that it could do his work and he could be home more.
Translation is important work and one that could help a lot of people. It's my hope that we get to the point where these models work entirely on locally carried resources.
I did several experiments recording from all the microphones I could on my iPhone and AirPods while out in the wild. My conclusion: it's impossible right now for that hardware given the microphones we have and what they pick up.
So much of what's spoken is at a combination of (a) high distance (b) low volume (c) background obscuration. Something that was clear as day to my ears would barely register on the mics. While context is of course an issue, the raw audio didn't have enough to even translate.
The one caveat is that there might be low-level (i.e., Apple-only) access to headphone microphones that capture the environment to do noise cancellation. I'm not sure though---I couldn't find them on any API.
For cases where you do have clear audio, existing apps (e.g., Google Translate) are so close to achieving this, but don't let you specify audio outputs with enough fine grained control. By default, it will start screaming out of your phone what you were attempting to silently translate.
So no matter what, conversations in your native speech would have to be delayed before translation.
That's a nice way of saying unemployed.
Everyone gets a personal tutor for hours a day.
I would absolutely love a VR game where I just need to work in China or Mexico all day and pick up the language that way.
He ended up rolling his own solution by standing up Whisper in one of our clusters and writing a basic front end and API to take his laptop’s mic input and chunk it every few seconds to send to the model and get back text in pseudo-realtime. We got him a pretty beefy Alienware so he wouldn’t be tied to the cluster GPUs. I can’t wait to see what he does with these new models!
I will note though that I feel safer getting an occasional bad word than I do having a translator straight up deceive me.
For example, "what the fuck" in English->Spanish is giving "qué diablos" output. Definitely toning down the meaning there.
If someone says something mean to me, I want to know it.
I told her then that the industry would be disrupted by AI before she retired.
Glad she pivoted. Really impressive results.
(quote from the video presentation in the link)
Just text to speech has gone too far. Audio books would be mainly generated on the fly like this?
I think some RPGs in some 5 years time might have something like this:
- A text file that outlines characters and a lose plot/Story line. Human written.
- 3D Mesh Generation based on character description via Transformers based models. Auto generated.
- Dialogues for each NPC via LLM.
- This TTS engine again based on such models.
Result - almost unlimited replayability. Or even edit text file, have a new world based on a new story line with characters having different personas.
It's not just where the verb is; sometimes I say something ambiguous, and my next utterance is supposed to acknowledge and remedy that. But if that ambiguity doesn't exist in the target language, I don't see how a simultaneous translator can convey the ambiguity, without knowing how the next utterance is going to refer to it.
Maybe that's why human simultaneous translators often seem to stumble or backtrack. I've never met someone whose job was simultaneous translation. It must be very difficult.
I'm impressed by this effort to convey non-linguistic elements of speech in translation. It's quite an achievement, and a very ambitious goal.
Aside: I wish I knew how speakers of tonal Chinese dialects express feeling, when tonality is supposed to convey semantics. When I hear chinese speakers, I can "hear" the feeling, but I don't know how they do it - it can't just be down to emphasis. (I learned some mandarin 50 years ago, at school. I learned the tones, but they didn't teach expression; and I was never taught by a native speaker, although there were language-lab tapes.)
It would be nice if they could layout what exactly is missing in terms of data to make a language work better, while the actual AI bit is out of reach for most of us maybe we could provide more data.
There is also a 60 sec limit and wonder if this is HuggingFace limitation or Seamless?
As a student in the mid-90s I worked on a system called Verbmobil at the German Research Center for AI and it did speech-to-speech for English, German and Japanese in very limited domain.
This was done via "classical" NLP: You had to model the domain with concepts, you needed sentence parsers, semantic engines, speech-to-text hand-crafted for 3 languages etc.
As it turns out, this approach is/was a dead-end.
I don't think this would fool anyone that I was a real native speaker of the target language, but for casual conversation this would work pretty much perfectly. It basically avoids all of the traditional pitfalls of machine translation, like the unnatural robotic voice that it outputs, the slow translation speed and huge latency for realtime conversation, and the loss of emotion.
So it can't be trusted at all then
Also: "This research demo is not open to residents of, or those accessing the demo from, the States of Illinois or Texas"
Besides the ACCEPTABLE_USE_POLICY, there's a CC BY-NC 4.0 (NonCommercial) license, a 'SEAMLESS_LICENSE' (NonCommercial), but also an MIT license? It would seem these other licenses contradict the MIT license, could somebody help clarify how these all interact in practice?
https://seamless.metademolab.com/expressive/?utm_source=meta...
It runs but any audio input (you will need to provide wav not mp3's) I tried (tried 20s/40s/300s) I get just one short sentence returned in target language that seems not related at all to my audio input (i.e. Tous les humains sont créés égaux).
Seems like some default text but it runs on full GPU for 10 minutes. Tons of bug reports in GitHub as well.
Text Translate works but not sure what is the context length of the model. Seems short at first glance (haven't looked into it).
Oh and why is Whisper a dependency? Seems not need if FB has their own model?
Edit: I can’t find the weights but if I’m reading the paper right anyone could train their own detector.
My ultimate goal is to have realtime translations of video conferences. I've moved to a new country, and while I'm super privileged that most of my colleagues speak English, we still have a number of "all hands" meetings that I get lost in pretty easily.
That would have been the knockout punch.
Especially because the head of AI at Meta is a French guy AFAIK (Yann Lecun).
What I would love to see is an ability to add my own voice (yes, at the risk of deepfakes) so that the model could "speak" in any language and sound more like me, not some random voice actor it was trained on.
Would allow more interactions of people that don’t speak the same language
Edit: by the upvote I guess it wasn't just me?
Attribution-NonCommercial 4.0 International
https://github.com/facebookresearch/seamless_communication/b...