I have hands-on experience in speech data annotation for AI training
I have hands-on experience in speech data annotation for AI training. I independently produced a bilingual Chinese-English short sentence comparison corpus as a portfolio project for an xAI audio trainer position. The work involved end-to-end speech data pipeline tasks: recording 10 phonetically diverse short sentences (covering declarative, interrogative, exclamatory, imperative, and negative sentence types in both languages) at professional-grade specifications (48kHz, 16-bit WAV), and performing dual-tier time-aligned annotation using Praat — a phoneme tier with 50 precisely segmented intervals (pinyin/English phonemes with boundary timestamps at millisecond resolution) and a word tier with 21 intervals mapping Chinese characters and English words to their corresponding audio segments. All silence intervals were systematically labeled. I documented the full recording environment parameters (noise floor ~25dB, mic distance 15-20cm, recording level -6dB to -3dB), identified and root-caused a Bluetooth audio latency issue (~500ms-1s delay) that interfered with annotation playback accuracy, and established a best-practice guideline to use wired headphones for phonetic annotation workflows. This project demonstrates practical proficiency in speech corpus construction, phoneme-level segmentation, bilingual alignment, annotation tooling (Praat/TextGrid), and quality control documentation.