Revision as of 18:50, 13 January 2023

Programming Etude #1: "Poets of Sound and Time"

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

In this programming etude, you are to write two programs using chuck with Word2Vec to help you create some experimental poetry involving text, sound, and time.

Due Date

• Etude #1: due Wednesday (1/18, 11:59pm) see "Deliverables" below
• Presentation: be prepared to share your work in class on Thursday (1/19)

Put the Disco in Discord

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

Things to Think With

• watch for ideas and inspiration: Allison Parrish's "Experimental Creative Writing with the Vectorized Word" (video)

Tools to Play With

• get the latest bleeding edge secret chuck build
  • macOS will install both command line chuck and the graphical IDE miniAudicle, and replace any previous ChucK installation.
  • Windows 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 will need to build from source, provided in the src
  • all platforms in order use console input (e.g., to run the sample code word2vec-prompt.ck), you will need to use command line chuck (console input is not available from within the miniAudicle IDE).
• NOTE: to return your chuck back to a pre-bleeding-edge state, you can always install the latest official ChucK release
• also, a Zoom recording of the ChucK tutorial from 2023.01.13

The Word Embedding Model

• lastly, you'll need to download three sets of pre-trained word vectors
  • glove-wiki-gigaword-50.txt -- pre-trained word vectors from Stanford GloVe (400,000 words x 50 dimensions)
  • glove-wiki-gigaword-50-pca-3.txt -- dimensionally reduced model using PCA (400,000 words x 3 dimensions)
  • glove-wiki-gigaword-50-tsne-2.txt -- dimensionally reduced model using t-SNE (400,000 words x 2 dimensions)
• each of these has their tradeoffs:
  • the 50D has the best similarity accuracy (i.e., tends to get better search results) -- but searching is more computationally intensive (and may introduce glitches in real-time audio when querying for similarity)
  • the PCA-3D version is a dimensionally-reduced (from 50 to 3) version of the above, making it much faster to do similarity retrieval (moreover the lower dimensionality in this case means it's feasible to use space partitioning techniques like KD-trees to significantly speed up the search); in addition to being real-time-audio friendly, the three dimensions can be readily mapped to control audio parameters (frequency, volume, rate, timbre, etc.) -- however, the quality of the similarity retrieval is noticeably weaker, as seemingly unrelated words will show up in the results.
  • the t-SNE-2D version, despite having one less dimension than the PCA-3D version, actually performs as well or better in terms of similarity compared to PCA-3D (partially due to the non-linear optimization of t-SNE compared with PCA's linear mapping of higher dimensions to lower dimensions); like 2D, 3D is friendly for mapping to audio parameters (and for visualization).
  • we recommend exploring word embeddings first using the 50D model, before moving on to the 3D and 2D versions for mapping and for real-time audio. It is also possible to first generate text, stored the generated text, and then render it and the audio in real-time. This approach effectively removes the real-time constraints by separating the processing into a generative stage and an efficient rendering stage; however, this also means that real-time interactions (e.g., to influence the generative processes) will be limited.

Sample Code

• free to incorporate these or use them as starting points
• you can find sample code here
  • word2vec-basic.ck -- basic example that...
    • loads a word vector
    • prints # of words and # of dimensions in the model
    • shows how to get a vector associated with a word (using getVector())
    • shows how to retrieve K most similar words using a word (using getSimilar())
    • shows how to retrieve K most similar words using a vector (using getSimilar())
    • uses the W2V helper class to evaluate an expression like "puppy - dog + cat" (using W2V.eval())
    • uses the W2V helper class to evaluate a logical analog like dog:puppy::cat:?? (using W2V.analog())
  • word2vec-prompt.ck -- interactive prompt word2vec explorer
    • this keeps a model loaded while allowing you to play with it
    • type help to get started
• example of poems
  • "Random Walk" -- a stream of unconsciousness poem of sound and time!
    • dependency: glove-wiki-gigaword-50-tsne-2
    • usage: chuck poem-randomwalk.ck or chuck poem-randomwalk.ck:america (to provide the starting word)
• we will be adding a few more examples over the weekend (but don't wait, start "practicing" your etude)

Express Yourself!

Using the ChucK/ChAI starter code for Word2Vec...

• write code to help you create some experimental poetry involving text, sound, and time.
  • text: use the Word2Vec object in ChucK and one of the datasets to help you generate some poetry
  • sound: use sound synthesis and map the words (e.g., using their vectors to control parameters such as pitch, volume, timbre, etc.) to sound
  • time: don't forget about time! make words appear when you want them to; synchronize words with sound; visually and sonically "rap" the words in time!
• create two poetic programs / works / performances:
  • make them as different as possible
  • for example one poem can be fully generated (you only need to run the chuck code, and the poem starts) and the other one interactive (incorporates input from the user, either through a a text prompt, or another means of input such as mouse / keypresses)

Some Prompts and Bad Ideas

• a poem can be about anything; hint: try starting with how you feel about a particular thing, person, or event
• starting with an existing poem and use word2vec to morph it over time
• an experimental love poem
• stream of consciousness
• remember "Jabberwocky" by Lewis Carroll? maybe your poem doesn't need to make sense to everyone
• HINT: try to take advantage of the medium: in addition to printing out text to a terminal (both a limitation and a creative constraint/opportunity) you have control over sound and time at your disposal
• HINT: experiment with the medium to embody your message -- for example, a poem about chaos where the words gradually become disjointed and nonsensical

Reflections

• write ~250 words of reflection on your etude. It can be about your process, or the product, or the medium, or anything else. For example, how did your poems make you feel? Did you show them to a friend? What was their reaction? What makes something "poetry" (versus, say, "prose")?

Deliverables

• create a CCRMA webpage for this etude
  • the URL should live at https://ccrma.stanford.edu/~YOURUSERID/356/etude1 or https://ccrma.stanford.edu/~YOURUSERID/470/etude1
  • alternately, you may use Medium or another publishing platform (but please still link to that page from your CCRMA webpage)
• your Etude #1 webpage is to include
  • a title and short description of the exercise (free free to link to this wiki page)
  • your poems in some form (this will depend on what you chose to do; since sound and time are involved, you could include a screen capture with sound)
  • your ChucK code
  • your 250-word reflection
  • any acknowledgements (people, code, or other things that helped you through this)
• submit to Canvas only your webpage URL