8ed76b40a6
Non-English voices drifted to English/wrong-language because the request's or profile's language wasn't reaching the model: - #533: generate_speech() read instruct/ref_text/seed from a resolved profile row but never row['language'] (and collapsed Auto→None), so a German archetype previewed in German yet generated in English on the user's own call (and via Docker/API). Fall back to the profile's stored language when the request didn't pin one; an explicit non-Auto request language still wins. (Frontend already sets the dropdown on profile-select; this is the authoritative backend fix.) - #505 (B2): the audiobook/longform synth hardcoded language=None, so the engine re-autodetected per chunk and a non-English clone flipped language mid-render. Add _resolve_default_language (request → profile → autodetect) and thread the resolved language through _build_synth/_prepare_synth/_render_longform_sse, the three longform request models, the preview path, and the resume manifest. Genuine Auto/unset behavior is unchanged. - #502 (partial): the duration estimator weights combining marks (U+0300–036F) at 0.0, so NFD/decomposed text under-allocated frames → rushed audio. NFC- normalize text at the estimator entry — fixes the whole diacritic-script class (no-op for precomposed text). (The residual "distorted" core still needs the reporter's sample; tracked separately.) Tests (fail-before/pass-after): profile language reaches the engine (German→de; explicit/Auto override semantics); longform synth gets the resolved language (→ja), not None; NFD vs NFC duration parity (Korean Hangul diverges ~3x pre-fix). Full suite: tests/ 1740 passed, backend/tests/ 114 passed. Co-authored-by: mergetest <test@local> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
85 lines
3.7 KiB
Python
85 lines
3.7 KiB
Python
"""Issue #502 (partial) — the duration estimator must be Unicode-normalization
|
||
-form independent.
|
||
|
||
``RuleDurationEstimator`` weights combining marks (U+0300–036F, category Mn) at
|
||
0.0. Decomposed (NFD) text — a base letter followed by combining tone/diacritic
|
||
marks, the canonical on-disk form for Vietnamese and many diacritic scripts —
|
||
therefore mis-allocates frames versus the precomposed (NFC) form of the *same*
|
||
text, producing rushed/garbled audio for the affected syllables.
|
||
|
||
The fix NFC-normalizes at the estimator's text entry point
|
||
(``calculate_total_weight``), so NFC and NFD inputs always yield the *same*
|
||
estimate. These tests fail before the fix (NFD diverges) and pass after.
|
||
"""
|
||
import unicodedata
|
||
|
||
import pytest
|
||
|
||
from omnivoice.utils.duration import RuleDurationEstimator
|
||
|
||
|
||
def _raw_weight(est, text):
|
||
"""Pre-fix weighting: sum per-char weights WITHOUT NFC normalization.
|
||
|
||
Calls the lru_cache-wrapped ``_get_char_weight`` via ``__wrapped__`` so the
|
||
fix's normalization step is bypassed — this reproduces the *old*
|
||
``calculate_total_weight`` behavior to prove the bug existed.
|
||
"""
|
||
return sum(est._get_char_weight.__wrapped__(est, c) for c in text)
|
||
|
||
|
||
# A single Korean Hangul syllable: precomposed (1 syllable block, weight 2.5)
|
||
# vs decomposed (lead/vowel/tail jamo). This is the clearest member of the
|
||
# normalization-divergence *class* — the bug is not Vietnamese-specific.
|
||
_KOREAN = "한" # 한
|
||
# A Vietnamese phrase exercising stacked tone + vowel-quality marks — the
|
||
# issue's named script. Stays stable across forms once normalized.
|
||
_VIETNAMESE = "Tiếng Việt nặng hỏi ngã sắc huyền"
|
||
|
||
|
||
def test_combining_marks_weigh_zero():
|
||
"""Guards the precondition the bug rests on: combining marks contribute 0.0,
|
||
so NFD text that splits a syllable into base+marks loses that weight."""
|
||
est = RuleDurationEstimator()
|
||
for cp in (0x0300, 0x0301, 0x0302, 0x0303, 0x0309, 0x0323): # Vietnamese tone marks
|
||
assert est._get_char_weight(chr(cp)) == 0.0
|
||
|
||
|
||
def test_nfd_diverges_from_nfc_before_normalization():
|
||
"""Fail-before guard: the OLD (un-normalized) weighting gives NFD a
|
||
different weight than NFC, which is exactly the defect."""
|
||
est = RuleDurationEstimator()
|
||
nfc = unicodedata.normalize("NFC", _KOREAN)
|
||
nfd = unicodedata.normalize("NFD", _KOREAN)
|
||
assert nfc != nfd # the syllable really does have distinct forms
|
||
assert _raw_weight(est, nfc) != _raw_weight(est, nfd)
|
||
|
||
|
||
@pytest.mark.parametrize("sample", [_KOREAN, _VIETNAMESE, "한국어 문장입니다"])
|
||
def test_calculate_total_weight_is_normalization_independent(sample):
|
||
"""After the fix, NFC and NFD inputs weigh identically (the fix normalizes
|
||
to NFC first). Non-zero, so the estimate stays meaningful."""
|
||
est = RuleDurationEstimator()
|
||
nfc = unicodedata.normalize("NFC", sample)
|
||
nfd = unicodedata.normalize("NFD", sample)
|
||
w_nfc = est.calculate_total_weight(nfc)
|
||
w_nfd = est.calculate_total_weight(nfd)
|
||
assert w_nfc == w_nfd
|
||
assert w_nfc > 0
|
||
|
||
|
||
@pytest.mark.parametrize("sample", [_KOREAN, _VIETNAMESE])
|
||
def test_estimate_duration_is_normalization_independent(sample):
|
||
"""End-to-end: the public estimate is identical for NFC vs NFD input.
|
||
|
||
Before the fix the Hangul case diverges ~3x (NFD over-counts as jamo);
|
||
after the fix both forms produce the same, correct estimate."""
|
||
est = RuleDurationEstimator()
|
||
ref_text, ref_dur = "Hello, world.", 1.5
|
||
nfc = unicodedata.normalize("NFC", sample)
|
||
nfd = unicodedata.normalize("NFD", sample)
|
||
est_nfc = est.estimate_duration(nfc, ref_text, ref_dur)
|
||
est_nfd = est.estimate_duration(nfd, ref_text, ref_dur)
|
||
assert est_nfc == est_nfd
|
||
assert est_nfc > 0
|