| 依据: 论文的核心”创新”是 transparency-first framework(per-module provenance flag)和将 ASR + diarization + signal analysis + subtitle export 整合到一个 pipeline 中。两者都不构成真实增量:(1) provenance/metadata tagging 是软件工程标准实践,在数据 lineage 系统(如 Apache Atlas, OpenLineage)、MLflow 等工具中早已实现;(2) 语音分析 pipeline 整合已被多个开源项目实现(free-search 发现 hadil-sgh/AudioPipelineTreatment “Real-Time Audio Intelligence |
Live speaker identification, speech-to-text, and sentiment analysis pipeline with Whisper & Wav2Vec2”、Tknika/azpidatzi “containerized audio transcription platform… Built with FastAPI”、snailbrainx/speaker-diarization-server “pyannote.audio + faster-whisper + emotion2vec + REST API + WebSocket streaming”)。论文中的每个技术组件都是对已有工具的直接调用:Whisper [1]、pyannote [2,3]、librosa [7]、VADER [8]、SHAP [17]、LIME [18]、Grad-CAM [20]、MC dropout [21]。论文没有提出任何新机制、新解释或问题重构。Table I 中的 “Proposed” 功能全部来自已有语音研究文献(jitter/shimmer [11], spectral features [7], loudness [10], turn-taking [13], emotion recognition [15]),论文只是将它们列入 roadmap,不属于本文的新颖性贡献。论文声称 “Less common in the literature is a single system that treats transcription, diarization, signal analysis, and workflow export as one pipeline with a shared provenance model”——但上述 GitHub 项目已经实现了类似的整合,且带有 provenance metadata。 |