Diversity Axes

A diversity axis defines what kind of diversity you want to measure by mapping to a specific embedding model. Different embedding models capture different aspects of text, so the same set of texts can have high semantic diversity but low stylistic diversity (or vice versa).

Most built-in axes embed text (a list of strings). A few embed a different modality instead — the speaker axis below embeds audio (a list of paths to audio files). The axis you pass decides how raw input is interpreted; the measure functions themselves are unchanged either way.

Built-in Axes

semantic (default)

Measures meaning-based diversity. Texts about different topics will score high; paraphrases will score low.

Default model: all-mpnet-base-v2
Alternatives: all-MiniLM-L6-v2 (faster, smaller)

from emb_diversity import log_determinant

# These are equivalent (semantic is the default)
log_determinant(texts)
log_determinant(texts, diversity_axis="semantic")

style

Measures writing style diversity. Texts written in different styles (formal vs. informal, simple vs. complex) will score high.

Default model: AnnaWegmann/Style-Embedding
Alternatives: StyleDistance/styledistance, rrivera1849/LUAR-MUD, AIDA-UPM/star

log_determinant(texts, diversity_axis="style")

speaker

Measures speaker diversity from audio. Pass a list of paths to audio files (not text) — each is embedded with a speaker-discriminative model where the same speaker’s utterances embed close together and different speakers embed far apart. Requires the audio extra (pip install emb-diversity[audio]).

Default model: speechbrain/spkrec-ecapa-voxceleb

audio_paths = ["alice_1.wav", "alice_2.wav", "bob_1.wav", "carol_1.wav"]
log_determinant(audio_paths, diversity_axis="speaker")

Using a Custom Model

You can bypass the axis system entirely by passing an embedding model directly:

log_determinant(texts, embedding_model="all-MiniLM-L6-v2")
log_determinant(texts, embedding_model="all-mpnet-base-v2")

The embedding_model parameter always overrides diversity_axis when both are provided.

Using Alternative Models

Each axis lists alternative models that work well for its purpose. To use an alternative, pass it via embedding_model:

# Use the faster alternative for semantic diversity
log_determinant(texts, embedding_model="all-MiniLM-L6-v2")

Discover available alternatives programmatically:

from emb_diversity import axes

for axis in axes.list_all():
    print(f"{axis.name}: default={axis.default_model}")
    if axis.alternative_models:
        print(f"  alternatives: {axis.alternative_models}")