Revision as of 02:51, 31 January 2023

Programming Project #2: "Featured Artist"

Music and AI (Music356/CS470) | Winter 2023 | by Ge Wang

In this programming project, we will learn to work with audio features for both supervised and unsupervised tasks. These include a real-time genre-classifier and a feature-based audio mosaic tool. Using the latter, create a feature-based musical statement or performance!

Due Dates

• Coding tutorial: Thursday evening (time TBD) | attendance is high recommended (will save you time later)
• Milestone (Phase One complete + Phase Two prototype): webpage due Wednesday (2/1, 11:59pm) | in-class critique Thursday (2/2)
• Final Deliverable: webpage due Wednesday (2/8, 11:59pm)
• In-class Presentation: Thursday (2/9)

Discord Is Our Friend

• direct any questions, rumination, outputs/interesting mistakes to our class Discord

Things to Think With

• read/skim the classic article "Musical Genre Classification of Audio Signals" (Tzanetakis and Cook, 2002)
  • don't worry about the details yet; first get a general sense what audio features and how they can be used

Tools to Play With

• get the latest bleeding edge secret chuck build (2023.01.23 or later!)
  • macOS this will install both command line chuck and the graphical IDE miniAudicle, and replace any previous ChucK installation.
  • Windows you will need to download and use the bleeding-edge command line chuck (for now, there is no bleeding-edge miniAudicle for Windows); can either use the default cmd command prompt, or might consider downloading a terminal emulator.
  • Linux you will need to build from source, provided in the linux directory
  • all platforms for this project, you will be using the command line version of chuck.
• NOTE: to return your chuck back to a pre-bleeding-edge state, you can always install the latest official ChucK release

GTZAN Dataset

• next, you'll need to download the GTZAN dataset
  • 1000 30-second music clips, labeled by humans into ten genre categories

Phase One: Extract, Classify, Validate

• understanding audio, audio features, FFT, feature extraction
• extract different sets of audio features from GTZAN dataset
• run real-time classifier using different feature sets
• run cross-validation to evaluate the quality of classifier based different features
• you can find relevant code here
  • start playing with these, and reading through these to get a sense of what the code is doing
  • example-centroid.ck -- a basic example of using ChucK's unit analyzer framework (things connected using the upchuck operator =^) to extract an audio feature:
    • generate an input (a 440hz sine wave) -- this can be any audio source, e.g., adc for the microphone
    • take a Fast Fourier Transform (FFT) on a frame of audio (size is determined by the FFT size)
    • using the output of FFT's analysis to compute the Spectral Centroid for that frame of audio
    • note how the ChucK timing is used to precisely control how often to do a frame of analysis
    • the .upchuck() is used to trigger an analysis, automatically cascading up the =^
  • example-mfcc.ck -- this is like the previous example, but now we compute a multi-dimensional feature, Mel Frequency Cepstral Coefficients (MFCCs)
  • feature-extract.ck -- in a "real-world" scenario, we would extract multiple features. a FeatureCollector is used to aggregate multiple features into a single vector (see comments in the file for more details)
  • genre-classify.ck -- using output of feature-extract.ck, do real-time classification by performing the same feature extraction and using k-NN to predict likelihood of each genre category (see comments in the file for more details)
  • x-validate.ck -- using output of feature-extract.ck, do cross-validation to get a sense of the classifier quality
• experiment by choosing different features and different number of features, extracting them on GTZAN, try the real-time classifier, and perform cross-validation
  • available features: Centroid, Flux, RMS, RollOff, ZeroX, MFCCs, Chroma, Kurtosis
  • try at least five different feature configurations and evaluate the resulting classifier using cross-validation
    • keep in mind that the baseline score is .1 (a random classifier for 10 genre), and 1 is max
    • how do different--and different numbers of--features affect the classification results?
    • in your experiment, what configuration yielded the highest score in cross-validation?
• briefly report on your experiments

Phase Two: Designing an Audio Mosaic Tool

• you can find phase 2 sample code here
• using what you've learned, build a database mapping sound frames (100::ms to 1::second) <=> feature vectors
  • curate your own set of audio files can be mixture of
    • songs or song snippets; we will perform feature extraction on audio windows from beginning to end; in essence each audio window is a short sound fragment with its own feature vector)
    • (optional) short sound effects (~1 second), you may wish to extract a single vector per sound effect
  • modify the feature-extract.ck code from Phase One to build your database of sound frames to feature vectors:
    • instead of generating one feature vector for the entire file, output a trajectory of audio windows and associated feature vectors
    • instead of outputting labels (e.g., "blues", "disco", etc.), output information to identify each audio window (e.g., filename and windowStartTime)
    • see reference implementation mosaic-extract.ck
  • note this does not require any labels, and like word2vec, we want to situate each sound window in a N-dimension feature space
• play with mosaic-similar.ck: a feature-based sound explorer to query your database and perform similarity retrieval (using KNN2)
• using your database and retrieval tool and concatenative synthesis and the mosaic-synth-mic.ck and mosaic-synth-doh.ck, design an interactive audio mosaic generator
  • feature-based
  • real-time
  • takes any audio input (mic or any unit generator)
  • can be used for expressive audio mosaic creation
• (optional) do this in the audiovisual domain; build a GUI for exploring sounds by similarity; will need to reduce dimensions (using PCA or another technique) to 3 or 2 in order to visualize

Phase Three: Make a Musical Mosaic!

• • use your prototype from Phase Two to create a feature-based musical mosaic in the form of a musical statement or performance
  • (optional) do this in the audiovisual domain

Reflections

• write ~300 words of reflection on your project. It can be about your process, or the product. What were the limitations (and how did you try to get around them?)

Deliverables

• create a CCRMA webpage for this etude
  • the URL should live at https://ccrma.stanford.edu/~YOURUSERID/356/hw2 or https://ccrma.stanford.edu/~YOURUSERID/470/hw2
  • alternately, you may use Medium or another publishing platform (but please still link to that page from your CCRMA webpage)
• your webpage is to include
  • a title and description of your project (free free to link to this wiki page)
  • all relevant chuck code from all three phases
    • phase 1: all code used (extraction, classification, validation)
    • phase 2: your mosaic generator, and database query/retrieval tool
    • phase 3: code used for your musical statement
  • video recording of your musical statement (please start early!)
  • your 300-word reflection
  • any acknowledgements (people, code, or other things that helped you through this)
• submit to Canvas only your webpage URL