SONIOX / DEEPGRAM RECOGNITION DECISION WORKSHEET
Free resource. Any evaluation usage needs authorization and an approved budget.
October8,2026. Guide: https://burki.dev/blog/soniox-vs-deepgram-streaming-recognition
No provider calls or model benchmarks were performed to populate this template.

1. DECISION AND PRODUCT AVAILABILITY
Business task / language / caller audio channel / reviewer:
Details whose errors would make the request unusable:
Provider/API account, permission, funding and recording policy:
Current Burki UI: Deepgram; AssemblyAI(BYO); Azure(BYO). No Soniox selector.
Soniox source adapter is not proof of an exposed workflow or successful invocation.

2. PUBLIC BENCHMARK RECEIPT
Source: https://github.com/pipecat-ai/stt-benchmark
Access date: October8,2026. Per-model run dates/geography not stated in summary.
Record the exact dataset/subset, judge and scoring revision from the original source:
Dataset / sample count / language filter / audio preparation:
Judge / scoring date / denominator rules / uncertainty:
Model | meanWER | pooledWER | transcript-return | TTFS median/P95/P99
stt-rt-v5 | ______ | ______ | ______ | ______
nova-3-general | ______ | ______ | ______ | ______
Record model-specific run dates/geography/settings when available; unknown stays unknown.
Keep original URLs and access dates; do not mix scoring revisions.
Proposed inference / evidence supporting it / evidence still missing:

3. FICTIONAL SUPPORT CASES
A. Correction: 'The serial is Q, L, seven ... sorry, Q, L, nine, two.'
   Intended latest identifier: QL92, subject to spoken confirmation.
B. Mid-identifier pause: 'Q, L ... nine, two.' Do not act on an incomplete prefix.
C. Unfamiliar product: 'It is a CedarArc unit.' Preserve uncertainty until confirmed.
D. No identifier available: 'I cannot find the label.' Do not invent a serial.
E. Handoff: 'I need to speak to staff.' Check configured handoff/availability honestly.
F. Ordinary question: 'What details should I have ready?' Do not force a known product.
These are proposed acceptance cases, not tested results or actual customer stories.

4. OWN EVALUATION RECORD (LEAVE BLANK UNTIL AUTHORIZED)
Case / exact audio reference / recording consent / date / attempt count:
Provider / explicit model ID / request ID / accepted request proof:
Language / endpoint / audio format / network / cold-or-reused connection:
Context or keyterms / endpoint settings / downstream instructions:
Final transcript / confirmed latest identifier / resulting assistant request:
Failure, missing transcript, premature finalization or unclear recording:
Reviewer decision / disagreement / unresolved field:
Compare minimal configurations first. Record special hints as separate configurations.
Hold audio and downstream instructions constant; log any unavoidable mismatch.

5. TIMING BOUNDARIES
t0 speech ends:
t1 final transcript received: (t1-t0 = defined recognition-finalization interval)
t2 response generation completes:
t3 audible output begins:
Record interruption/pause behavior separately. Do not infer full-call timing from TTFS.
Distribution/attempts/interval/failed attempts:

6. COST TEMPLATE
Soniox Oct8 listed streaming units:
input audio USD2/1Mtokens; input/output text USD4/1Mtokens.
Audio tokens ___ * 2/1000000 = USD___
Input text tokens ___ * 4/1000000 = USD___
Output text tokens ___ * 4/1000000 = USD___
Vendor ~USD0.12/hour is an illustrative usage estimate, not a flat tariff.
Deepgram Oct8 listed promotional PAYG Nova3 mono USD0.0048/min:
Minutes ___ * .0048 = USD___
Multilingual: verify separate .0058/min rate and exact language set.
Optional keyterms minutes ___ * .0013 = USD___
Account terms/allowances/tax/other add-ons where applicable:
Reasoning/synthesis/telephony/platform costs separately:
Sources: https://soniox.com/pricing ; https://deepgram.com/pricing

7. CURRENT BURKI NOVA3 CONFIGURATION TO REVIEW
Assistants -> Voice -> expand Model and provider choices
Conversation engine: Standard pipeline · advanced provider choices
Within that panel: Advanced speech recognition -> Deepgram
General model ID nova-3-general, intended recognition language.
Optional Keywords & key terms -> Key terms; record list as another configuration.
Review readiness and Usage & billing before any saved/budgeted session.
Current live editor inspected read-only: no Soniox option, no save/call.
No Soniox UI switch or accepted provider result is demonstrated by this guide.

8. DECISION RECORD
Acceptance criteria chosen before review:
Observed outcomes and number of attempts:
Costs based on actual billing units:
Product path and permissions available:
Chosen next step / owner / recheck trigger / date:
Unknowns remain unknown; source selection and a published benchmark are not acceptance.