VOXTRAL MINI / SMALL TRANSCRIBE 3: OFFLINE REVIEW WORKSHEET
Original free resource | Prepared October 25, 2026

This is an unperformed paired-review design. No recording is supplied, no file
was uploaded and no paid transcription was run for this resource. Provider and
production costs are separate from this free worksheet.

1. DEFINE THE RECORDED-AUDIO TASK
Business question:
Recording origin and permission:
Language and qualified reviewer:
Required output and downstream use:
Critical facts whose loss would change the task:
Who resolves an unclear recording before a business action?

An offline file review does not establish live conversational timing. A live
recognition provider selector is not a saved-file retranscription integration.

2. FREEZE THE INPUT AND REFERENCE
Local file / controlled source:
Unchanged source version / hash:
Actual duration and format:
Reference transcript, checked against what was actually spoken:
Reference reviewer / revision / date:
Normalization rule, if calculating WER:
Business-meaning acceptance rule, kept separate from word scoring:
Approved processing scope / account owner / allowed spending:

If a speaker departs from the script, correct the reference to the actual
recording. Do not grade against the intended script without checking it.

3. ORIGINAL FICTIONAL FIXTURE
Business: Alder Print
Speaker script, not an existing recording:

This is about project Q seventeen. Sorry, Q seventy, not seventeen. We need
four display cards, not fourteen. Please ask Nila to check the revised proof
before anyone prints it.

Verbatim reference: NOT CREATED; requires an approved recording and review.
Expected current interpretation:
- Project Q70; Q17 is the corrected earlier detail.
- Quantity four; fourteen is superseded.
- Nila should check the revised proof before printing.
- No printing, notification or project update has been performed.

Do not delete the spoken correction from the reference merely because the
team needs one final project code in its separate task summary.

4. THREE ADDITIONAL ORIGINAL CASES - ALL UNPERFORMED
V2 - Preserve the negative instruction
Script: Use the matte proof for review. Do not print from the glossy draft.
Expected meaning: Matte proof is for review; glossy draft is not print authority.
Observation: NOT RUN

V3 - Uncertain proper name
Prepare an approved recording with a real task-relevant name that the test
speaker is authorized to use. Keep its spelling/reference separately reviewed.
Expected: Inspect actual name errors; do not assume plausible spelling correct.
Observation: NOT RUN

V4 - Same words, changed recording conditions
If later authorized, retain a separately identified version with different
recording conditions. Do not replace the clean source without recording it.
Expected: Compare each model on the same chosen version, not different inputs.
Observation: NOT RUN

5. MATCH THE REQUESTS
Model A requested: voxtral-mini-transcribe-3-0
Model B requested: voxtral-small-transcribe-3-0
Same input route and exact file version:
Same language and other options:
Required timestamp/output settings:
Vocabulary settings, if any, explicitly recorded:
Account arrangement and current tariff evidence:
Number of planned attempts:
Order / changed conditions recorded:

Keep model aliases out of an unverified version claim. Record requested ID and
returned metadata. No provider acceptance is claimed by these planning fields.

6. PER-OUTPUT RECORD - COPY FOR EACH MODEL / FIXTURE / ATTEMPT
Model requested / returned metadata:
Fixture / recording hash / reference revision:
Request/response reference (no secrets):
Complete output or failure:
Actual accepted usage / cost evidence:
Words omitted, inserted or substituted:
Q70 correction retained? Evidence:
Four rather than fourteen retained? Evidence:
Name and before-printing instruction retained? Evidence:
Different spelling/format with same meaning:
Human review time and unresolved questions:
Decision: ACCEPT FOR THIS CASE / REVISE / UNRESOLVED

Keep raw outputs. An edited task summary is not the provider's original
transcript, and a completed transcript is not an accepted business action.

7. REFERENCE SCORE AND TASK DECISION
If calculating WER, name the reference/normalizer and keep its source version.
Do not compare a hand-cleaned candidate with an unedited candidate and call
that a fair model result. Keep critical corrections separately visible.

Actual result table, initially empty:
Model / fixture / attempts / word score method / critical fact errors /
unresolved cases / review time / actual usage / remaining unknowns

No scores are supplied. The published aggregate benchmark discussed in the
article is context, not a measured result for this worksheet or language.

8. COST SCENARIO - NOT AN INVOICE
Current standard provider rate used, with access date:
Planned audio minutes:
Mini provider transcription subtotal:
Small provider transcription subtotal:
Review/reprocessing work, separately measured:
Other storage/network/platform/taxes/account terms:
Allowed total and decision owner:

For the article's hypothetical 1,000 audio minutes, listed standard tariffs
produce USD3.50 Mini or USD4.00 Small, a USD0.50 difference before other costs.
Confirm the applicable arrangement; do not infer credits or a free paid run.

9. CHOOSE THE ACCEPTED SCOPE
Cases where both outputs meet the chosen task:
Cases where one remains unreliable:
What a vocabulary change would require reassessing:
What a different language/route/source needs separately:
Decision and supporting evidence:
Human follow-up owner / next review trigger:

No native Burki Mistral transcription connector is established here. Preserve
existing Burki transcripts separately from any independently approved provider
evaluation. No API calls, installs or product settings were performed.

Companion guide:
https://burki.dev/blog/voxtral-mini-vs-small-transcribe-3
Current model and first-party tariff references:
https://docs.mistral.ai/models/voxtral-mini-transcribe-3-0
https://docs.mistral.ai/models/voxtral-small-transcribe-3-0
https://docs.mistral.ai/inference/pricing