Plain-Language Acoustic Feedback for Indian L2 English: A CPU-Only Pipeline with Bottom-Up Vowel-Nucleus Segmentation

We describe a CPU-only acoustic pipeline that produces plain-language pronunciation feedback for Indian second-language (L2) English speakers preparing for IELTS and TOEFL Speaking assessments. The system runs on a single 2-core virtual machine without GPU acceleration, processes a 90-second monologue in approximately eight minutes, and emits both a structured per-finding acoustic profile and a templated natural-language explanation readable by parents and learners at lower IELTS band levels. Two contributions distinguish the design. First, we replace standard wav2vec2-CTC forced alignment — whose 20 ms frame quantization is too coarse for vowel-length measurement — with a bottom-up vowel-nucleus segmenter derived from Praat pitch and intensity contours; the segmenter splits raw voiced runs at intensity local minima and is bounded by a per-nucleus duration cap. Second, every acoustic finding is paired with a templated natural-language summary at the engine layer, so that downstream consumers (web client, PDF report, voice-assistant interface) need not own the linguistics translation. We report preliminary engine-side measurements on a pilot cohort of seven IELTS Speaking samples drawn from two Indian L1 backgrounds (Hindi, Bengali), and describe the N=60 examiner-agreement protocol whose blind human grading will be reported in follow-up work. The pipeline is consistent with the published Indian L2 English phonetics literature (Yarra et al., 2019; Jain et al., 2021). This is a defensive publication: it discloses the system architecture publicly to establish prior art, while withholding numeric coefficients, thresholds, and verbatim template strings that constitute the production system's trade secrets. Withheld values are flagged as [withheld] in the body.