Japanese version with cloned voice

Gemini helps to visualize

Then helps to understand pronunciation for lip motion

Some preprocessing in Audacity and CLI magic with a video note and we have a fake

A way to convert datasets from JSONL to DuckDB quickly.

#duckdb

That example shows this dataset: https://huggingface.co/datasets/speech-uk/generated-news-comments

The same method was used to create a table for https://huggingface.co/datasets/speech-uk/voiced-news-comments dataset

DuckDB gives ability to fastly prototype

In the Speech-UK initiative these two datasets will be used to crowd-source data for Ukrainian ASR

Next step: create a Telegram bot using https://github.com/AmarnathCJD/gogram to make voice crowd-sourcing available to beta users

Hmm…

When I implement tasks in my Speech-UK initiative (btw, follow it here - https://huggingface.co/speech-uk) some projects arise like this one.

It gives you ability to filter out bad samples from a dataset based on another model Meta made recently.

https://github.com/egorsmkv/audiobox-aesthetics-inference

#ml #audio

Also, added a Gradio app to https://github.com/egorsmkv/marblenet-inference

You can perform VAD operation on large files (if you have a GPU) with ease now

#audio

NSA at home:

They don't have a prebuilt package for Ubuntu 24, so I am building it from the source code

It has a lot of dependencies, btw