Audio Classification using Convolutional Neural Net

Define audio environments for the Ahmed keyword by organizing raw data into silent, in-car with family, and outside-noise scenarios, and treat unlisted environments as positive data for the neural net.

Record the silent car environment for audio classification with a convolutional neural net, capturing ten to fifteen minutes of silent cues like ac, radio, glass, seatbelt, and the keyword Ahmed.

Record the car's second environment with family in the car, capturing conversations and in-cabin noises like seat belt and handbrake, without dummy audio; aim for ten to fifteen minutes.

Record the third real car environment, family talk inside the car with the glasses down, with outside traffic sounds and occasional silence, aiming for 10 to 15 minutes.

Demonstrates converting mobile audio recordings from m4a to wav for machine learning, using cloud convert and basic file organization, with optional Python automation.

compose positive audio by injecting the keyword Ahmed into a negative background, record clips with a screen capture tool, convert to wave, and test the resulting positive file.

Develop audio classification workflows by recording from a car family talk environment, capturing positives and negatives, converting webm to wav, and organizing files for environment-based dataset preparation.

Slice audio into 3-second clips for neural network classification, using Google Colab and Google Drive, with librosa and soundfile to name and save each clip.

Finalize slicing the audio into 61 three-second clips, balancing silent negative and silent positive classes for neural net training.

Create a train and test folder split for audio clips, allocating about 90% to training and 10% to testing, while balancing across environments to evaluate the model.

Demonstrates a Python-based train-test split for the eclipse audio dataset, balancing by class with 90 percent training and 10 percent testing, using Colab, os, and shutil.

Implement a Python driven train-test split by copying or moving negative and positive clips into train and test folders using a for loop and shuttle commands, with counters tracking totals.

Sort audio clips into negative and positive classes in a Colab workflow, mount Google Drive, copy files into a train sorted directory, and count 165 negatives and 165 positives.

Load and play audio files in Python by sorting train sorted files into negative (label zero) and positive (label one), then load with librosa and play using IPython display.

Explore four stages to convert audio clips into a spectrogram. Load audio with librosa, transform to frequency domain, apply amplitude to decibel, and expand to 3d tensor for cnn.

Process train sorted WAV clips through five stages: load audio with Librosa, apply Fourier transform, convert to decibel spectrogram, resize to a 3D tensor, and assemble into audio_clips for TensorFlow.

Learn labeling basics for audio classification by pairing each audio clip with a corresponding label, using tensor slices and batching in TensorFlow to train a neural net.

Label audio clips as positive or negative by creating and concatenating numpy labels, then attach them to audio data with tensor slices to form a labeled dataset for training.

Learn batching as a data-pipelining strategy to feed a convolutional neural network: group two audio clips with two labels per batch, inspect tensor shapes, and prepare the dataset for training.

Learn to assemble a sequential CNN with convolutional and max pooling layers, flatten and dense layers, then compile with Adam and binary cross-entropy, and review the model summary.

Train cnn for audio classification by creating a tensor data set, splitting with take and skip (160 train, 4 test), and fit the model to reduce loss and improve accuracy.

The lecture demonstrates testing a cnn model for audio classification by processing test clips, predicting positive or negative labels, and validating with loss, accuracy, and multiple audio samples.

Save the trained cnn as an h5 file, set the target directory for Raspberry Pi deployment, load the saved model, and run predictions to validate performance on the dataset.

Set up a Raspberry Pi 5 with 8 GB RAM and a micro SD card, and boot the operating system; install an active cooler with thermal pads and fit a case.

Learn to assemble a Raspberry Pi 5 by mounting thermal pads on the processor, RAM, and IO controller, installing an active cooler, wiring the power, and testing before OS installation.

Install the Raspberry Pi five operating system with Raspberry Pi Imager, selecting the 64-bit image and a micro SD card, customize hostname and wifi, enable ssh, then write and verify.

Connect to the Raspberry Pi 5 via SSH, update and upgrade packages, enable the VNC server with raspi-config, and access the pi remotely with the VNC viewer, while monitoring temperature.

Set up a raspberry pi 5 environment, connect via ssh and vnc, create the Ahmed audio folder, and create and activate a python 3 virtual environment to install tensorflow.

install and test librosa on Raspberry Pi 5, download and move an h5 model to the device via FileZilla, then load the model in app.py using Keras.

Record three-second audio clips on a Raspberry Pi 5 using a USB microphone, save them as WAV files in the project folder, and feed each clip to a convolutional neural net to determine whether it contains the word Ahmed.

Record a three-second audio clip with a USB mic on Raspberry Pi 5, preprocess with librosa, FFT, and decibel spectrogram, then predict with h5 model and drive hardware via GPIO.

Raspberry Pi 5 uses gpio pins to drive a micro servo motor, moving a plastic arm when the AI model detects Ahmed.

Test a convolutional neural network-based audio classification model on Raspberry Pi 5, using a mic and servo to act when predictions exceed 0.9, even in noise.

Explore audio classification in Python using cnn and rnn on groans clips from ICU dataset, balancing positives and negatives, with training in Google Colab and Google Drive integration.

Learn to prepare an audio dataset for convolutional neural nets by loading audio with Librosa, applying short-time Fourier transform, converting to decibel spectrograms, and expanding dimensions for neural network input.

Learn to label audio as positive or negative, enforce fixed input shapes, generate labels with numpy, and build a shuffled, batched TensorFlow dataset for CNN training.

Build and train a cnn for audio classification using keras, with train/test split, pooling, dropout, and a sigmoid binary output.

the lecture demonstrates adapting a cnn-based audio classification workflow to an rnn with lstm using keras, detailing architecture changes, activation choices, and binary classification performance.