ELEVEN V4 VS V4 TURBO COMPARISON WORKSHEET
Burki free resource | Checked October 7, 2026
Original home: https://burki.dev/blog/eleven-v4-vs-v4-turbo

No benchmark has been performed for this resource. Blank records are UNTESTED.
Provider calls and synthesis can cost money. Arrange a bounded authorized test
budget before running cases. Use fictional data and an agreed recording policy.

1. DEFINE THE DECISION BEFORE TESTING
Workload (short assistant replies / produced narration / other): _____________
Language / listener population: _____________
Must-pass meaning/delivery cases: _____________
Maximum acceptable caller-to-audio delay and precise definition: _____________
Budget / approved owner: _____________

2. FIX THE SHARED CONFIGURATION
Model A: eleven_v4 | Model B: eleven_v4_turbo
Provider account/plan/access review: _____________
Assistant draft: _____________  Standard pipeline: _____________
Same accessible voice ID for both models: _____________
Language: _____________ (start fixed; Automatic is a separate test)
Recognition model/settings: _____________
Reasoning model/settings and approved business facts: _____________
Response text revision: _____________
Actual provider settings/transport/output format if visible: _____________
Network/connection conditions and location: _____________
Cold-start or reused connection: _____________
Voice/format incompatibility prevents identical setup? _____________

CURRENT UI LIMITS
Voice -> Conversation engine: Standard pipeline · advanced provider choices
-> Speech settings -> Advanced voice stack -> Voice provider: ElevenLabs
-> exact Voice model.
V4 Voice tuning shows delivery guidance and Language choices. It hides legacy
sliders; do not invent Stability/Similarity adjustments in the current form.
Provider v4 docs support Stability/Similarity, not Style/Speed/SSML.
GPT Live separate provider choice controls Announcement voice only.

3. REUSABLE FICTIONAL SCRIPTS
A. "We open at nine in the morning. I can explain the enquiry process."
B. "Do you mean Wednesday morning or Thursday morning?"
C. "I cannot confirm a booking from this enquiry. Staff need to review the request."
D. "I can explain the available staff contact options."
E. Longer approved answer containing an ordinary name, time and service limit.
F. Plain response versus the same response with one permitted delivery tag.
   Keep tag trial separate from the plain baseline and identical across models.
G. Authorized interruption case: caller corrects the question after reply begins.
H. A language-specific case approved by a reviewer, after the fixed-language run.
Keep actual task/action outcomes separate from synthesis quality.

4. MARK TIMESTAMP BOUNDARIES
t0 = caller actually stops speaking (chosen observation method): _____________
t1 = synthesis work/request submitted (chosen receipt method): _____________
t2 = audible assistant output begins (chosen observation method): _____________
t3 = requested speech completes / is interrupted: _____________

Whole interaction delay = t2 - t0.
Observed synthesis-plus-delivery interval = t2 - t1.
Neither automatically equals a provider's model inference time. If timestamps
are unavailable, write UNMEASURED instead of guessing from a recording feeling.
Hold timestamp methods, output format and conditions constant across the pair.

5. CASE / ATTEMPT RECORD (copy for repeated attempts)
Case ID / attempt / model / date: _____________
Exact text / voice ID / language: _____________
Cold or reused connection: _____________
t0 / t1 / t2 / t3 and units: _____________
Request accepted / rejected / no receipt: _____________
Actual playback / unavailable: _____________
Error / interruption / retry: _____________
Meaning understood by reviewer: YES / NO / UNTESTED
Missed or altered words: _____________
Delivery suitability, with agreed scale: _____________
Model-blind reviewer ID and listening order: _____________
Business task/action result (separate): _____________

6. SUMMARIZE WITHOUT CHERRY-PICKING
Attempts per model/condition: _____________
Errors/retries/interruptions counted: _____________
Median measured interval and units: _____________
Percentile only if justified by sample count/method: _____________
Blind preference counts / ties / reviewer count: _____________
Unmeasured or non-comparable fields: _____________
Do not hide failures or call a small clip preference a production accuracy rate.

7. COST SNAPSHOT
Provider pricing date / URL / plan / expiry: _____________
Unit rate and billable character definition: _____________
Characters in proposed workload: _____________
Synthesis-only estimate = characters / 1000 * unit rate.
Separate platform, recognition, reasoning, telephony, taxes and plan allowances.
October7 advertised rates through October12: v4 USD.022/1000chars,
Turbo USD.011/1000chars; displayed ordinary rates USD.08 and USD.04.
Recheck before use. These are a dated provider advertisement, not Burki quotes.

8. DECISION
Choose / revise / defer: _____________
Evidence and important tradeoff: _____________
Must-pass failures still unresolved: _____________
Retest cases and review date: _____________
Do not infer a head-to-head winner from vendor Turbo-only latency figures.

Voice selection: use Assistant voice. For Saved custom voice, expand
Custom voice identifier and enter the actual permitted Custom voice ID.