Spaces:
Running on Zero
Running on Zero
test(01-06): add failing tests for the vowel table and banker's frame quantisation
Browse files- 13-entry VOWEL_TO_VISEME including the uppercase devoiced A I U E O
- to_frame() proven round-half-to-even against 8 exact-half vectors
- tests/test_visemes.py +101 -0
tests/test_visemes.py
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""AVTR-02 and VOIC-03: the viseme timeline builder.
|
| 2 |
+
|
| 3 |
+
This is the densest test file in Phase 1 because ``visemes.py`` is the one function whose
|
| 4 |
+
failure modes are all silent. A wrong rounding mode, a lowercase-only vowel table or a
|
| 5 |
+
float-accumulating loop each produce a timeline that looks entirely plausible in a debugger and
|
| 6 |
+
visibly wrong on the avatar's face.
|
| 7 |
+
|
| 8 |
+
Every fixture here was captured from a real ``voicevox_core`` 0.17.0 synthesis by
|
| 9 |
+
``tests/fixtures/make_synth_fixtures.py``; every duration compared against is read from the WAV
|
| 10 |
+
header, never summed from the query.
|
| 11 |
+
|
| 12 |
+
Nothing in this file imports ``voicevox_core`` - the builder is a pure function over plain dicts,
|
| 13 |
+
so the quick loop stays installable-free and fast.
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
from __future__ import annotations
|
| 17 |
+
|
| 18 |
+
import math
|
| 19 |
+
|
| 20 |
+
import pytest
|
| 21 |
+
|
| 22 |
+
from japanese_avatar.voice.visemes import (
|
| 23 |
+
DEVOICED,
|
| 24 |
+
FRAMERATE,
|
| 25 |
+
VOWEL_TO_VISEME,
|
| 26 |
+
to_frame,
|
| 27 |
+
viseme_for,
|
| 28 |
+
)
|
| 29 |
+
|
| 30 |
+
# frames -> to_frame(frames / FRAMERATE). Round-half-up would give the third column.
|
| 31 |
+
# Verified to round-trip exactly through f / 93.75 * 93.75 on CPython.
|
| 32 |
+
BANKERS_VECTORS = [
|
| 33 |
+
(0.5, 0, 1),
|
| 34 |
+
(1.5, 2, 2),
|
| 35 |
+
(2.5, 2, 3),
|
| 36 |
+
(4.5, 4, 5),
|
| 37 |
+
(5.5, 6, 6),
|
| 38 |
+
(10.5, 10, 11),
|
| 39 |
+
(20.5, 20, 21),
|
| 40 |
+
(100.5, 100, 101),
|
| 41 |
+
]
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def test_vowel_mapping():
|
| 45 |
+
"""AVTR-02: every VOICEVOX vowel symbol, including the devoiced uppercase ones."""
|
| 46 |
+
assert VOWEL_TO_VISEME["a"] == "aa"
|
| 47 |
+
assert VOWEL_TO_VISEME["i"] == "ih"
|
| 48 |
+
assert VOWEL_TO_VISEME["u"] == "ou"
|
| 49 |
+
assert VOWEL_TO_VISEME["e"] == "ee"
|
| 50 |
+
assert VOWEL_TO_VISEME["o"] == "oh"
|
| 51 |
+
|
| 52 |
+
# Japanese devoices /i/ and /u/ between voiceless consonants constantly - です is
|
| 53 |
+
# "d e s U", した is "sh I t a". VOICEVOX emits those as UPPERCASE. A lowercase-only
|
| 54 |
+
# table freezes the mouth on every polite form.
|
| 55 |
+
for v in "AIUEO":
|
| 56 |
+
assert VOWEL_TO_VISEME[v] == VOWEL_TO_VISEME[v.lower()], v
|
| 57 |
+
assert DEVOICED == frozenset("AIUEO")
|
| 58 |
+
|
| 59 |
+
assert VOWEL_TO_VISEME["N"] == "closed" # ん
|
| 60 |
+
assert VOWEL_TO_VISEME["cl"] == "closed" # っ
|
| 61 |
+
assert VOWEL_TO_VISEME["pau"] == "closed" # silence
|
| 62 |
+
|
| 63 |
+
# Exactly 13: an extra key is an invented symbol, a missing key is a frozen mouth.
|
| 64 |
+
assert len(VOWEL_TO_VISEME) == 13, sorted(VOWEL_TO_VISEME)
|
| 65 |
+
|
| 66 |
+
# Devoiced vowels still open the mouth, just less.
|
| 67 |
+
assert viseme_for("u") == ("ou", 1.0)
|
| 68 |
+
assert viseme_for("U") == ("ou", 0.5)
|
| 69 |
+
assert viseme_for("I") == ("ih", 0.5)
|
| 70 |
+
assert viseme_for("N") == ("closed", 0.0)
|
| 71 |
+
assert viseme_for("pau") == ("closed", 0.0)
|
| 72 |
+
|
| 73 |
+
# A new VOICEVOX symbol must fail loudly rather than animate wrongly.
|
| 74 |
+
with pytest.raises(KeyError):
|
| 75 |
+
viseme_for("x")
|
| 76 |
+
with pytest.raises(KeyError):
|
| 77 |
+
VOWEL_TO_VISEME["q"]
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
@pytest.mark.parametrize(("frames", "expected", "half_up"), BANKERS_VECTORS)
|
| 81 |
+
def test_frame_quantisation_bankers(frames, expected, half_up):
|
| 82 |
+
"""AVTR-02: round-half-to-even, matching np.round inside VOICEVOX itself.
|
| 83 |
+
|
| 84 |
+
The VOICEVOX source carries the warning 「NOTE: `round` は偶数丸め。移植時に取扱い注意。」
|
| 85 |
+
Python's built-in round() is banker's, so a Python port matches for free; JavaScript's
|
| 86 |
+
Math.round() is round-half-up and would disagree on every exact-half boundary. That is an
|
| 87 |
+
independent, concrete reason this arithmetic lives in Python.
|
| 88 |
+
"""
|
| 89 |
+
assert FRAMERATE == 93.75
|
| 90 |
+
assert FRAMERATE == 24000 / 256
|
| 91 |
+
|
| 92 |
+
seconds = frames / FRAMERATE
|
| 93 |
+
# Fail loudly if a future Python stops round-tripping, rather than passing by luck.
|
| 94 |
+
assert seconds * FRAMERATE == frames, (frames, seconds * FRAMERATE)
|
| 95 |
+
|
| 96 |
+
assert to_frame(seconds) == expected
|
| 97 |
+
|
| 98 |
+
# Provably NOT round-half-up: on the exact-half boundaries the two modes disagree.
|
| 99 |
+
if expected != half_up:
|
| 100 |
+
assert to_frame(seconds) != math.floor(seconds * FRAMERATE + 0.5)
|
| 101 |
+
assert math.floor(seconds * FRAMERATE + 0.5) == half_up
|