What problem does it solve? Streamed LLM-to-speech pipelines fail in subtle ways: choppy clause chunking, split words from sub-word tokens, leaked control tags, spliced audio, and runaway reverb feedback. This Skill documents five real defect patterns found in a production voice pipeline and the exact fixes for each. ## Core Features & Use Cases - Clause-boundary chunking guidance: Explains why timer-based flushes and off-by-epsilon threshold comparisons break semantic segmentation, and how to verify fixes against measured latency. - Fragment and token handling: Covers filtering degenerate punctuation-only fragments before synthesis, merging rather than dropping them, and reconstructing whole words from sub-word LLM token streams with two-sided boundary checks. - Audio correctness patterns: Details incremental control-token parsing with partial-token hold-back, drop-newest backpressure for PCM queues, and correct OLA crossfade and feedback-delay reverb implementation. - Use Case: When a voice agent's speech output sounds choppy, drops words, or clips unexpectedly, use this Skill to identify which of the five documented defect classes matches the symptom and apply the corresponding fix. ## Quick Start Ask the assistant to review the TTS chunker and audio queue code for the streaming speech defects described in this Skill.
