TTS INTERRUPTION EVALUATION WORKSHEET
Free resource | Checked October 26, 2026
Companion: https://burki.dev/blog/flux-tts-vs-sonic-3-6-interruptions
Original evaluation plan. No paid synthesis or phone test was performed.
A worksheet does not grant API access or make usage free.

1. PIN THE TEST ARRANGEMENT
Date / timezone:
Reviewer:
Authorization for any later paid test:
Provider and exact resolved model, if known:
Voice:
Region / endpoint:
Output encoding and sample rate:
Playback application and device:
Telephony path, if tested:
Direct API vs Burki native capability evidence:
Unknown values must stay unknown.

2. ORIGINAL ADMINISTRATIVE FIXTURE
Fictional Westbridge reception.
Proposed speech: I have noted Tuesday afternoon. The team will confirm availability.
Caller correction: Actually, Wednesday morning.
Expected business status: request for staff review, not booked appointment.
Proposed next speech: I have changed the request to Wednesday morning. The team
will confirm availability.
Condition before saying it: verify actual saved request, or explain unresolved
status. A synthesis event is not a calendar or business-system receipt.

3. APPLICATION LABELS
Old local generation: reply-17
New local generation: reply-18
Labels are application-owned, not additional provider fields.

At interruption:
Local old-output suppression time:
Local playback position / how measured:
Provider control send time:
Provider acknowledgment or warning evidence:
Late old chunks received:
Late old chunks dropped:
Any old chunk played after stop:
Proposed old text retained separately:
Evidence of delivered old text:
Caller correction retained:
Business save result or unresolved status:

Provider reminders:
Flux offset is session-wide, not per-turn; an advancing position is required.
Its text split requires playback_offset. Stop local audio before reconciliation.
https://developers.deepgram.com/docs/flux-tts/interrupt-handling

Cartesia context guide: cancel stops not-yet-generating requests; current
generation continues. Retain local stale-output handling.
https://docs.cartesia.ai/use-the-api/tts-websocket/contexts
General websocket wording differs; this worksheet does not resolve it by test.

4. OFFLINE DESIGN REVIEW FIRST
A. Inject a late chunk labeled reply-17 after application switches to reply-18.
Expected: chunk is not eligible for playback.
Observed:
Pass / fail / unknown:

B. Inject a missing provider outcome.
Expected: unknown outcome is retained; no full-delivery claim is invented.
Observed:
Pass / fail / unknown:

C. Review the dialog history after an interruption mid-date.
Expected: proposed speech is not automatically recorded as fully delivered.
Observed:
Pass / fail / unknown:

D. Inject business save failure after caller correction.
Expected: next speech does not claim the changed request was saved.
Observed:
Pass / fail / unknown:

5. OPTIONAL LATER AUTHORIZED AUDIO REVIEW
Keep the script, cue, output format and playback path comparable.
Do not run paid tests simply because this worksheet exists.

Case / voice / cue position / observed old audio after cue / history result:
Near beginning:
During Tuesday afternoon:
At sentence boundary:
Corrected callback number:

Distinguish clocks:
- Application saw interruption.
- Provider received control / application received reply.
- Local device stopped old output.
- Far-end listener stopped hearing old output, if actually observed.
Do not fill an unobserved clock with another clock's measurement.

6. QUALITY AND COST DECISION
Independent benchmark basis: article's dated English stock-voice preference
snapshot. It is not an interruption or multilingual benchmark.
Reviewer comprehension notes for this actual script / voice / language:
Actual character/credit usage if later measured:
Billing plan and access date:
Application failure to repair:
Provider behavior needing clarification:
Native product support needing verification:
Decision / owner / next evidence:

Keep generated, queued, locally played, heard and saved as separate outcomes.