Latency, Memory, and Verifiable Spoken-Format Constraints in Local ASR–LLM–TTS Pipelines: An Open-Domain Factorial Evaluation

A spoken reply requires a host-side automatic speech recognition (ASR), language model (LLM), and text-to-speech (TTS) pipeline. Latency, device memory, and spoken-format adherence are properties of that pipeline. We evaluate it on 999 English read-aloud HeySQuAD questions under two recognizer regimes. Each turn is question-only: the LLM receives the recognized question text and a fixed system prompt, without the SQuAD passage. Scoring is text-side on archived replies plus synthesis timing; no listening test or conversational usefulness rating is claimed. Conversational latency is speech-end-anchored time-to-first-audio (TTFA). Format pass is the strict conjunction of programmatic hard checks on the short-opener prompt; the five-word opener dominates failures. It is not a general instruction-following or task-achievement score. On a measured validation subset under streaming Vosk, TTFA medians were 1.28–1.54 s (2.03 s on a 27B cell); endpoint delay dominated the wait except on that cell. Absolute TTFA is not reported for the buffered Whisper quality arm. Corpus word error rates were 0.200 on streaming Vosk and 0.049 on Whisper. Four-bit quantization reduced device memory by a factor of about two to three. Under the full conjunction, Q4 raised format pass on Llama, showed no detected change on Qwen, and lowered it on Gemma, and majority pass occurred only at 27–32B. Without the five-word opener check, both effects change: all twelve temperature-0 Vosk cells pass on 88.8–99.6% of items, and the Llama Q4 gain and Gemma Q4 loss are gone. Smaller models could satisfy the full conjunction while showing low coverage of the intended question across all items. Short-answer reference-span presence remained low under the question-only prompt. Under the higher-error regime, intended-question coverage fell more than format pass.