Spaces:
Sleeping
Sleeping
Commit History
10 days ago
-
fix(pronunciation): improve phoneme alignment and audio preprocessing pipeline
- Replace unconstrained DTW with global alignment and explicit gap penalties to avoid false deletions - Normalize phoneme symbols between eSpeak and wav2vec2, expand IPA features table - Reward length-only vowel mismatches (e.g., /i/ vs /iː/) with partial credit - Fix stress index mapping and prevent spurious stress errors - Improve audio preprocessing with edge padding and safer VAD - Deduplicate wav2vec2 model inference in api endpoint - Add automated alignment test suitecb82b68
19 days ago
-
feat(pronunciation): enrich rule feedback with articulatory guidance
f570a96 -
feat(pronunciation): support US and UK accents in IPA evaluation
6fa6232 -
fix(pronunciation): replace AutoTokenizer with Wav2Vec2CTCTokenizer and add phonemizer
a8f54e4
22 days ago
-
feat(SCRUM-34): 7-component pronunciation assessment
- Add POST /grade/pronunciation/practice endpoint - Implement DTW phoneme alignment with IPA feature distance (place/manner/voicing weighted: 0.5/0.35/0.15) - Add pronunciation_analysis/ package with 7 components: 1. individual_sounds: Wav2Vec2 + DTW alignment 2. word_stress: Parselmouth F0 per-syllable + librosa onset detection 3. sentence_stress: F0 peak mapping + isochrony rhythm check 4. linking: Rule-based C+V / gemination / intrusive / elision 5. intonation: F0 contour rising/falling/flat detection 6. chunking: Pause detection + speech rate WPM analysis 7. contracted_words: 64-entry contraction dictionary - Add weighted scoring by target_type: word (60/35/5%), phrase (35/20/25/10/10%), sentence (30/20/15/15/10/5/5%) - Add rate limiting: 30 req/min per IP sliding window - Add praat-parselmouth + dtw-python dependencies - Add 20 unit tests (all passing, mocked AI models) Closes SCRUM-34ca71644
-
Deploy V15 Clean (Removed binary files history)
d939bae