1107 languages but no Czech or Slovak? Many languages with way fewer speakers made it to the list. I wonder what we did to Meta...
no chinese TTS either.
Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
221–230 of 232 posts
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#222Earlier quoted context omitted.
You can finetune whisper, stable diffusion, and LLM up to about 15B parameters with 24GB VRAM. Which leads you to what hardware to get. Best bang for the $ right now is definitely a used 3090 at ~$700. If you want more than 24GB vram just rent the hardware as it will be cheaper. If you're not willing to drop $700 don't buy anything just rent. I have had decent luck with vast.ai
There is the world of used Nvidia Teslas, like the M40. Very cheap, but some assembly required.
I own a P100, P4, M40. They either lack ram or speed. Also, unless you are putting them in a server you have to cool them.
If your goal is to learn ML don't tinker with very obsolete hardware. Rent or buy something modern.
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#223Great, now the Terminators will be barking orders at each other in languages I can't understand.
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#224Earlier quoted context omitted.
I wrote a colab to run TTS, it takes some time to install dependencies : https://colab.research.google.com/drive/11tZ1c7kQJ5uKImm9t-C... Overall I got pretty poor results in english and french, I guess it would require some fine-tuning
Thanks for this! Out of curiosity, how does vits play into this? I know nothing about the repo. Also, I used the following to play the sound after it is generated: from IPython.display import Audio, display display(Audio(output_file, autoplay=True))
I think vits is just a repo that implements the TTS model
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#225Earlier quoted context omitted.
Speaking as someone who has encountered similar difficulties, this response has strong 'Draw the rest of the owl' vibes
Speaking as someone who has solved these difficulties hundreds of times, "draw the rest of the owl" doesn't tell you the specific things to google to get detailed examples and tutorials on how millions of others have sidestepped these repeated issues.
[dead]
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#226Earlier quoted context omitted.
Wait why is renting a GPU in the cloud not a solution? You can even try multiple options and see which ones are capable enough for your use case. Look into some barebones cloud GPU services, for example Lambda Labs which is significantly cheaper than AWS/GCP but offers basically nothing besides the machine with a GPU. You could even try something like Vast in which people rent out their personal GPU machines for chea…
I have not seen many setups that wouldn’t pay itself back (including energy in my case) within a year (sometimes even 6 months) with buying vs renting. For something that pays itself back that fast, and that is without renting it out myself, just training with it, I cannot see how I would want to rent one. Edit; on Lambda labs, the only exception seems to be the H100; it would be 1.5 years or so, but even 2 years wou…
I think the downside to buying hardware as well is that compared to other tech, this LLM/ML stuff is moving very quickly, people are great at quantising now (whereas I only really saw it done before for the coral edge TPUs etc).
Someone could buy an H100 to run the biggest and bestest stuff right now, but we could find that a model gets shrunk down to run on a consumer card within a year or two with equivalent performance.
I suppose it makes sense if someone wants to be on the bleeding edge all the time.
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#227Imagine if we used these types of models for like 500 years and it locked their vocabulary in time, disallowing any further language blending; then somehow the servers turned off and nobody could communicate across language barriers anymore. Someone should write that down in some sort of short-story involving a really tall structure.
When a jet skier accidentally runs into a tree that's called a snow crash, aha
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#228Earlier quoted context omitted.
Nah, full-on babelfish is simply not possible. The meaning of the beginning of a sentence can be modified retroactively by the end of the sentence. This means that the Babelfish must either be greatly delayed or awkwardly correct itself every once in a while.
I wonder if an esolang exists that could be a universal target though (if a language can handle any ambiguity by appending, then it could always be output without backtracking).
Ithkuil?
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#229Wonder how this compares with deepgram offering. has anyone used/tried/compared or even read enough literature to compare. The WER rates showed in deepgram are still better than the largest MMS and the specific use case based fine-tuned models (zoom meetings, financial calls etc) probably make a bigger difference. WDYT ?
Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages
#230Earlier quoted context omitted.
Imagine if someone said this about dictionaries.
Imagine if your dictionary was constantly updated to modify your behavior
That statement makes no sense.
My point was that languages evolve, even though we write books and make dictionaries with fixed-in-time vocabulary lists. Similarly, langauges will evolve even if LLMs, for mysterious reasons, have a fixed-in-time vocabulary. I was responding to "and it locked their vocabulary in time, disallowing any further language blending".