Retain accent? Appolio

[torn comet]

I have 40 minutes of data of my clean audio, I am training with batch size 4, till 400 epoch but i am loosing my accent if I try to speech to speech using AI TTS speech, any solution?

[opal sand]

your dataset may be mixed up with some other voices or roleplayed voices, also it may be overtrained (pls check the tensorboard graph). use batch size 8 or 10 minutes/less recommended for batch 4.

[torn comet]

I am using colab and the dataset has only my recorded cleaned voice

[opal sand]

you still haven't answered about tensorboard and overtraining

besides the accent incompatibility to the TTS, overtraining might possibly cause that issue

[torn comet]

yes this, I am using colab notebook and it is showing a blank where it shoudl supposed to show the tensorboard and after few hours it got disconnected so I am now using kaggle and setting up things

how can I setup the tensorboard so I can share the results?

[opal sand]

the kaggle one has everything you need, read the guide https://docs.applio.org/applio/getting-started/other-alternatives#kaggle

[torn comet]

reading this and running the training

still no tensor board stats thing on it

[opal sand]

wdym? again there should be some tfevents file in logs/yourmodel/eval

also try in chrome/edge browser and make sure the tensorboard page is not blocked

and try start over training

[torn comet]

okay, i will try it out

[torn comet]

hey finally getting things, Please check if you have time, the training is still ongoing:
https://100d-34-122-73-226.ngrok-free.app/?darkMode=true

[opal sand]

well it's still 100-ish epochs, but remember to test some checkpoint

[torn comet]

at what epoch it starts working as expected?

300 to 400?

[opal sand]

try, try, and...

https://tenor.com/view/lsfc-try-aberdeen-standard-investments-gif-15801775

[torn comet]

oki

[torn comet]

and yeah, when do I have to extract the index file? after the training, right?

[scarlet edge]

index file can be done before or after, doesnt matter

[opal sand]

no it can't be extracted

[torn comet]

i have created an index file before the training and its size is 389MB and it is showing me an error when I try to load it like this (to be safe I have downloaded the project files from kaggle if they can help).

[opal sand]

wrong place, it is for inference

[torn comet]

so the index file I need for inference should be dropped here or the index file I got from somewhere else?

got this error in the console:

Traceback (most recent call last):
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/gradio/queueing.py", line 624, in process_events
    response = await route_utils.call_process_api(
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/gradio/route_utils.py", line 323, in call_process_api
    output = await app.get_blocks().process_api(
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/gradio/blocks.py", line 2015, in process_api
    result = await self.call_function(
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/gradio/blocks.py", line 1562, in call_function
    prediction = await anyio.to_thread.run_sync(  # type: ignore
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/anyio/to_thread.py", line 56, in run_sync
    return await get_async_backend().run_sync_in_worker_thread(
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/anyio/_backends/_asyncio.py", line 2505, in run_sync_in_worker_thread
    return await future
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/anyio/_backends/_asyncio.py", line 1005, in run
    result = context.run(func, *args)
  File "/kaggle/tmp/.venv/lib/python3.10/site-packages/gradio/utils.py", line 865, in wrapper
    response = f(*args, **kwargs)
  File "/kaggle/working/program_ml/tabs/download/download.py", line 46, in save_drop_model
    file_name.split("_nprobe_1_")[1].split("_v2")[0].split(".index")[0]
IndexError: list index out of range

btw I trained for like 500 epochs and the quality of the model is shit, I guess have to refine the dataset

[torn comet]

Hey, I want to ask few things:
is using the titan pre-trained model will improve the results?(Any other suggestion you want to give?)
what is the minimum and maximum duration of the dataset is a must, like what if I am of the minimum side like 4 minutes? what settings like batchsize is good? and what about 1 hour?

[fast dust]

Dataset

A set of audio files compressed into a .zip file, used by RVC for voice training. The quality & length of the dataset are the biggest determining factors of the final quality of the model.

[torn comet]

the link is not working

[torn comet]

and Its been like 10 times and as soon as i try to load the index file I made I got an error, only for the index I made if I use an index from someone else I got no error while loading it