This web application visualizes the phonetic/acoustic similarity between words as an
interactive 2D edge-weighted graph, an acoustic network embedding (ANE) visualization.
Key Components:
- Acoustic Embeddings: Utilizes Woojay Jeon's (@Apple)
Acoustic Neighbor Embeddings (ACN)
to map words into a high-dimensional space where similar-sounding words are closer together.
There is a setting at the top of the page for the dimensionality of the global embedding.
- t-SNE Projection: Applies t-Distributed Stochastic Neighbor Embedding (t-SNE)
to reduce
the high-dimensional embeddings into a 2D layout. We get the best results with a large global
embedding,
choose the neighborhood words, and then reduce to 2D for display.
- Distance Layout (optional): Choose "Distance (radial)" in the Layout menu
to put the query word in the centre and each neighbor at its true embedding distance, with
dashed rings marking round distances. It shows exact distances to the query and works best
with few neighbors; t-SNE is better for seeing which neighbors sound like each other.
- Interactive Graph: Renders query words and their nearest acoustic neighbors
using D3.js. Hover over nodes to inspect individual similarity distances.
By Kent Slaney and Malcolm Slaney with an implementation by Antigravity.
For more details see this
CCRMA blog post.
The source code is available
on GitHub.