nice

spamtons never gonna sound right is he

you could try finetuning a hifi-gan model

i have no idea how to do that

alright ill test it out

which one is it?

oh wait this one?

yeah

if im transcribing someone in another language as arpa but they have some english lines should i just make the whole list arpa

You can write the english lines in english or arpa, doesn't really matter

alright thanks

Text to video but with long length (2 minutes)

hmhm https://github.com/jianfch/stable-ts

also they added a thing to accommodate prompt engineering

oh this is really nice

gonna have to upgrade my install because i tried prompt engineering on polish with

prefix

and it failed catastrophically

though with polish it seems to veer off anyway, like 5% of cases had random other languages at the end or sometimes from the start even

Do we know if it's possible to train a sound diffusion model yet o(-( I've seen Mr. Pikachu send some results a while ago and it intrigues me to this day

I just really am curious about AI dreaming ideas in my music style

Ok now, I'm trying to download the PyTorch as said in the uberduck tutorial but appearently 10.2 and 11.3 have been discontinued what's the best subsitute

Or they aren't working on my computer

anyone?

currently getting speaker embeddings for all the tacotron voices, using this mess of code

py
api_key = # not giving you mine
data_path = "/content/drive/MyDrive/uberduck-embeddings.json"

!pip install speechbrain
!pip install tinydb

import os
import numpy
import requests
import shutil
import torchaudio
import traceback

from speechbrain.pretrained import EncoderClassifier
from tinydb import TinyDB, Query
from tinydb.storages import JSONStorage
from tinydb.middlewares import CachingMiddleware
from tqdm.auto import tqdm

classifier = EncoderClassifier.from_hparams(source = "speechbrain/spkrec-xvect-voxceleb", savedir = "pretrained_models/spkrec-xvect-voxceleb")
q = Query()

headers = {
    "accept": "application/json",
    "authorization": f"Bearer {api_key}"
}

voice_list = requests.get("https://api.uberduck.ai/voices?mode=tts-basic", headers = headers).json()

with TinyDB(data_path, storage = CachingMiddleware(JSONStorage)) as db:
  for i in tqdm([i for i in voice_list if not db.search(q.voicemodel_uuid == i["voicemodel_uuid"])]):
    try:
      r = requests.get(f"https://api.uberduck.ai/voices/{i['voicemodel_uuid']}/samples").json()
      audio = [torchaudio.load(requests.get(j["url"], stream = True).raw)[0] for j in r]
      embeddings = [classifier.encode_batch(j).squeeze().tolist() for j in audio]
      db.insert({**i, "embeddings": embeddings, "avg_embedding": numpy.average(embeddings, axis = 0).tolist()})
    except Exception:
      print(f"fucky wucky at {i['name']}")
      traceback.print_exc()

oh cool it seems that the for loop is being very slow to initialize now

fixed it seems

as my voice is on uberduck i'll record a new clip of myself speaking some random language and use that as a first test case

what the hay

usecases: - moderators (me!) looking for duplicate models - users looking for soundalikes of blacklisted voices - viewers curious about which voice meme zee used this time