A Implementation of SpecAugment with Tensorflow & Pytorch, introduced by Google Brain

SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition

Tensor2tensor experiment with SpecAugment

End-to-end speech recognition on AISHELL dataset.

fast SpecAugmentation code with numpy and scipy

tf 2.0 implementation of Listen, attend and spell

A minimalistic Tensorflow 2.x Keras layer which applies SpecAugment to its input

Emotion recognition with IEMOCAP datasets. We compare the results with SpecAugmentation and CodecAugmentation. For audio codec implementation, we have selected opus.

XSpeech: A Novel Deep Learning Approach to Classifying Stutters

End-to-end English speech recognition in PyTorch from scratch: CNN + BiLSTM + CTC trained on 100h LibriSpeech. 22.6% WER greedy, 12.2% with beam search + 4-gram LM under 7 GPU-hours on a single laptop GPU.

Radio signal spectrogram classification with PyTorch, SpecAugment-based data augmentation, and a pretrained EfficientNet-B0 model for four-class signal recognition.

Speech recognition toolkit featuring SpecAugment, Whisper fine-tuning, LAS ASR, synthetic data generation, benchmarking, and a desktop GUI. Built with PyTorch, Transformers, Librosa, and CustomTkinter.

PyTorch implementation of Transformer-based Automatic Speech Recognition with attention mechanisms, SpecAugment, CTC loss, and mixed precision training. Achieves competitive WER/CER on LibriSpeech.

Simple numpy-based implementation of SpecAugment

Performs data augmentation as according to the SpecAugment paper. Modified from Lingvo (TensorFlow > 1.10.0).

An Audio Classification task with two types of inputs to the CNN models for intended work using Tensorflow.

REST API based on PyTorch (ResNet18) for classifying 50 categories of natural and household sounds (rain, chainsaw, glass breaking, etc.) from audio files. Mel spectrograms + FastAPI. Val accuracy 86%. Trained in Google Colab on ESC-50.

FastAPI service for music genre classification (GTZAN, 10 classes) using a CNN with Mel spectrograms. 78% test accuracy, SpecAugment data augmentation, fully offline inference.