AutoEITAudio-to-text transcription for secondadditional language learner data

The Problem General-purpose Automatic Speech Recognition (ASR) often fails language researchers and learners because it is designed to "clean up" speech. In second-language learner data, it is critical to capture exact production, including disfluencies, attempts at pronunciation, and interjections. Current models often struggle to segment these recordings correctly or erroneously "fix" a student's grammatical and phonological errors. The Solution My proposal implements a robust, three-phase automated pipeline: - Pre-processing: I will implement noise reduction using tools like noisereduce or DeepFilterNet and utilize industry-standard VAD (Voice Activity Detection) such as Silero VAD or Pyannote.audio to segment audio into single-sentence blocks. - Transcription: Using high-fidelity models like faster-whisper, the system will capture all vocalizations with word-level timestamps. This phase focuses on preserving the learner's actual speech patterns rather than idealized grammar. - Post-processing: The pipeline will identify pauses, match transcriptions to target sentences, and use Large Language Models (LLMs) with Constrained Prompting. This allows the system to distinguish between intelligible learner errors (to be preserved) and actual [gibberish] (to be flagged). Key Deliverables - Python-based CLI Tool: An efficient command-line interface capable of transcribing raw MP3 files. - High Accuracy: A system designed to achieve >90% agreement with human transcribers. - Structured Output: Final results formatted as an Excel sheet for easy analysis by researchers and learners.