Thanks — I hear you: one sentence is intermittently skipped by ElevenLabs TTS in production and it’s hard to reproduce. I can help investigate—first, a few quick questions and a short set of checks that often find the cause. Could you share: • The exact sentence (or a masked version if sensitive). • Whether SSML is used for that sentence and whether

enable_ssml_parsing

is true. • Your ElevenLabs TTS settings:

use_tts_aligned_transcript

(true/false),

chunk_length_schedule

values, and whether preemptive generation is enabled. • Whether this happens when the agent is speaking (possible interruption) or only when starting speech. • Any relevant logs around the event (TTS errors, cancellations, or streaming flushes) and approximate timestamps. While you gather that, please try these checks: • Disable SSML parsing for that sentence (or validate/strip SSML tags) to rule out malformed SSML causing skips. • Temporarily set

use_tts_aligned_transcript=False

to see if alignment logic is altering the input. • Increase

chunk_length_schedule

entries (or adjust boundaries) so the sentence isn’t split at edge cases during streaming. • Confirm interruption/cancellation logic: ensure no inbound audio or cancel call coincides with the sentence being queued. • Search logs for TTS errors or cancellation events at the timestamps when the sentence was skipped. If you share the sentence and the settings/log snippets, I’ll analyze them and suggest the next targeted step (e.g., config change, fallback behavior, or a reproducible test). Sources: Text and transcriptions | LiveKit Documentation | ElevenLabs TTS | LiveKit Documentation | Sequential Pipeline Architecture for Voice Agents | LiveKit, with interactive elements