voxpreference / backend /tests /test_stability.py
thebickersteth's picture
Preference Grammar pipeline: one image for frontend and backend
9b3bb37 verified
Raw History Blame Contribute Delete
1.69 kB
from pg.ot import Candidate, Violation
from pg.pipeline import Token, preferred_stability, variant_shares
from pg.tenets import Settings, load_typology, load_units
def tok(speaker, variant):
return Token("r.wav", speaker, "think", 0, 1, variant, 1.0, {}, "", "")
def think_attested():
one = lambda: Violation(1) # noqa: E731
return [Candidate("tɪŋk", "t", violations={"DisaM": one()}),
Candidate("θɪŋk", "θ", violations={"Pho-S": one()}),
Candidate("sɪŋk", "s", violations={"I-Ac": one(), "DisaM": one()})]
def test_speaker_weighting_stops_one_speaker_dominating():
tokens = [tok("a", "t")] * 9 + [tok("b", "θ"), tok("c", "θ")]
_, by_token = variant_shares(tokens, "tokens")
_, by_speaker = variant_shares(tokens, "speakers")
assert round(by_token["t"], 2) == 0.82 and round(by_speaker["t"], 2) == 0.33
def test_stability_is_high_when_theta_is_clearly_frequent():
tokens = [tok(f"s{i}", "θ" if i % 2 else "t") for i in range(20)]
res = preferred_stability(tokens, think_attested(), load_units()["TH-voiceless"], load_typology(), "I-Ac >> ContR >> FreQ >> DisaM >> Pho-S >> Ped-C", Settings(), n=300)
assert res["preferred_rate"]["θ"] > 0.95
def test_stability_reveals_a_knife_edge_near_the_frequency_threshold():
# θ used by about 20% of speakers: whether it passes FreQ depends on who is sampled
tokens = [tok(f"s{i}", "θ" if i < 4 else "t") for i in range(20)]
res = preferred_stability(tokens, think_attested(), load_units()["TH-voiceless"], load_typology(), "I-Ac >> ContR >> FreQ >> DisaM >> Pho-S >> Ped-C", Settings(), n=300)
assert 0.2 < res["preferred_rate"].get("θ", 0) < 0.8