Reliability of Speech Recognition for Conversational Human–Robot Interaction: A Multi-Corpus Analysis of Denoising, Voice Activity Detection, and Transcription

Rising numbers of isolated nursing home residents and nursing staff shortages motivate voice-based conversational assistants for monitoring and companionship; both require reliable transcription without an enrolled speaker profile to reduce staff burden and serve as a baseline for speaker-based methods. We benchmarked a cascaded denoising, voice activity detection (VAD) and speech-to-text pipeline (DeepFilterNet3, Silero VAD, faster-whisper large-v3) at library defaults on five English corpora: utterance-based files (LibriSpeech, AMI clips on digital silence, LibriMix mixtures) and real recordings (CHiME-6). We report the pooled-count word error rate (WER) with cluster-bootstrap intervals, offline VAD F1 and embedding-based semantic error rate (SER). The best WER ranged from 4% on clean read speech to 68% on the CHiME-6 close-talk channel, whose reference omits other participants’ speech. External=VAD routing lowered WER in all 10 matched utterance-based comparisons (0.1 to 5.3 points); on CHiME-6 the effect depended on channel and reference (without denoising: close talk −65.3 points against the wearer’s reference, +12.4 against the all-participants reference; distant array +10.4). Denoising helped substantially only on the CHiME-6 close-talk channel. VAD F1 showed no consistent association with WER on utterance-based files (r = +0.11). The SER misses fluent fabrication. The findings suggest a provisional, in-sample rule needing held-out validation, including older-adult speech.