Released package to convert audio folders to torch datasets of spectrograms, labels and masks, useful for training.

GitHub Repo

Released JaVAD: Just Another Voice Activity Detector, SOTA vad open source package that is faster than every existing open source package I tested it against.

GitHub Repo

JANET (Directional Alignment) -based diarization package v0.1.1 New lite model (lite_v2) provides 20% improve in Diarization Error Rate

GitHub Repo

Simple way to visualize JADIA results.

pip install jadia-plot

JANET (Directional Alignment) -based diarization package v0.1 Uses K-means, best at diarizing dialogues with two people.

GitHub Repo

Tired of dealing with TQDM logging issues, created a TQDM for CometML and W&B. It detects if W&B or Comet ML module is up and replaces TQDM logger with a custom logger with large periods between updates.

GitHub repo

Simple Whisper trainer, directly compatible with OpenAI repo (no need for Huggingface libraries)

GitHub repo

Repo for custom Directional Alignment layer created to better extract features from speech spectrograms, compared with Conv1d/Conv2d Resnet-like stacks.

GitHub repo

Custom implementation of George Hinton's Forward-Forward algorithm with rather stupid but working loss.

GitHub repo

FastAPI/uvcorn custom microservice to process PDF and extract information using ChatGPT

GitHub repo

Custom Alignment layer to improve extraction capabilities.

Link

Correct implementation of Wide ResNets (compared to other found on github).

Top1 Cifar10 for WRN 28x10 dropout=0.3