Researchconformerctcspeech recognitionpronunciation evaluation
Developer Trains Conformer CTC Pronunciation Grader
|
7.1
A developer trained a Conformer encoder with CTC loss on ~300 hours (AISHELL-1 and Primewords) to grade Mandarin pronunciation, using SpecAugment and Viterbi forced alignment. Models ranged from 75M to 9M parameters, with tone accuracy ~98.3% and TER ~4.8–5.3%; INT8 quantization produced an ~11MB browser-deployable demo. The system uses Pinyin+tone tokens and blank-frame filtering to improve scoring.
Scoring Rationale
Practical, reproducible on-device ASR demo with empirical metrics; limited by single-source personal project and non-peer-reviewed results.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
