INWORLD TTS-2 / TTS-2 FLASH STEERING EVALUATION WORKSHEET
Prepared October 11, 2026. Free worksheet; voice generation is metered.

This is an unperformed protocol with fictional scripts, not benchmark results.
Neither TTS-2 model ID is exposed in the verified native Burki selector.
No voice generation, paid test or provider action is performed by this file.

1. DEFINE THE REQUIREMENT BEFORE SCORING
Business message owner: ______________ Reviewer: ___________
Required outcome: __________________________________________
Is documented steering essential? __________________________
Is unsteered speech acceptable? ____________________________
Script approval / version: _________________________________
Caller languages / accent expectations / local names: _______
Permitted voice and permission evidence: ____________________
Direct provider access and budget owner: ___________________

Eligibility decision:
- TTS-2 supports steering; Flash ignores steering instructions.
- Do not give Flash a steering success score for pleasant neutral speech.
- A supported control does not guarantee your desired listening result.

2. FIX THE COMPARISON CONTRACT
Exact models: inworld-tts-2 / inworld-tts-2-flash
API route and SDK version: _________________________________
Voice identifier for each model: ___________________________
Spoken language: ___________________________________________
Delivery mechanism: [ ] inline tags [ ] request instruction
Instruction mechanism supported by chosen transport? ________
Encoding / sample rate / client region / playback: __________
Connection reused or newly opened: _________________________
Number of planned attempts, including failures: _____________
Fixed segmentation and context strategy: ___________________
Request version and actual metadata location: ______________
Budget ceiling and stop condition: _________________________

3. SCRIPT A: NEUTRAL BASELINE, SAME SPOKEN WORDS
The entrance is beside the library, not the loading gate.
What would you like the team to clarify?

Evaluate both models only after access, permission and budget are resolved.
Listener should correctly repeat the entrance and understand the final question.

4. SCRIPT B: DIRECTED PASSAGE WITH EXPLICIT RESET
[say slowly with clear articulation] The entrance is beside the library,
not the loading gate. [reset] What would you like the team to clarify?

Steering eligibility: TTS-2 only. Intended behavior, not measured result:
- entrance instruction deliberately articulated;
- question no longer depends on the earlier slow instruction;
- all spoken words retained and intelligible.

5. SCRIPT C: SCOPE CONTROL, SAME WORDS WITHOUT RESET
[say slowly with clear articulation] The entrance is beside the library,
not the loading gate. What would you like the team to clarify?

Compare B and C without changing model, voice, text or audio format.
In a separate WebSocket exercise, put the final question in a later message
within the SAME context. Keep reset/no-reset variants separately versioned.
A new message is not automatically a new instruction scope.
These are planned checks, not promised output or an installed SDK example.

6. OBSERVATION RECORD: ONE ROW PER ATTEMPT
Blind label: ______ Exact model: ______ Script/version: ______
Request ID / retained evidence: _____________________________
Success/error and reason: ___________________________________
All words retained? ______ Entrance repeated correctly? ______
Unwanted delivery carryover? _______________________________
Question intelligible? ______ Voice consistent? _____________
Listener preference, kept separate from correctness: ________
First byte boundary / measured time: _______________________
First audible speech boundary / measured time: ______________
Last audible word boundary / measured time: _________________
Actual billable characters / charge: ________________________
Reviewer / language / date: _________________________________

7. AGGREGATE WITHOUT INVENTING A WINNER
Sample count including errors: _____________________________
Accepted messages / failed messages: ________________________
Missing words / misunderstood directions: ___________________
Timing boundary, unit, measurement location and percentile: __
Provider claims kept separate from your own measurements? ___
Total actual usage / accepted output: _______________________
Evaluation covers only English fixture, or other languages? _
Other accents/local names require new acceptance evidence: __

Vendor P90 first-byte figures are server-side and exclude network. They are
not measured caller-heard latency. No shared independent latest-pair quality
benchmark was established in the guide. Do not fill result fields from ads.

8. COST AND DEPLOYMENT DECISION
Plan / currency / rate access date: _________________________
Current applicable character rate / units: __________________
Billable character-count receipt: __________________________
Retry, application, LLM, storage, platform and phone costs: __
Accepted-output cost and calculation: _______________________
Chosen workflow and evidence supporting it: _________________
[ ] Documented eligibility checked, runtime still untested
[ ] Authorized evaluation completed with retained evidence
[ ] Hold deployment because acceptance failed or is incomplete

Primary references:
https://docs.inworld.ai/tts/capabilities/steering
https://docs.inworld.ai/tts/tts-models
https://docs.inworld.ai/tts/synthesize-speech-websocket
https://inworld.ai/pricing
Original guide:
https://burki.dev/blog/inworld-tts-2-vs-tts-2-flash