AZURE SPEECH VS DEEPGRAM NOVA-3: RECOGNITION DECISION WORKSHEET
Reviewed October 12, 2026. Free resource; production recognition is billed.
Original guide: https://burki.dev/blog/azure-speech-vs-deepgram-nova-3

This is an unperformed evaluation plan. It does not contain Burki benchmark
results. Provider model-family selection is not evidence of an accepted call.
Azure's native identifier azure-speech is not a verified LLM Speech 2607 revision.

1. DEFINE ONE TASK
Example: community workshop session request, with staff confirming availability.
Capture a preferred date, caller name and callback number; never invent a booking.
Your task: _________________________________________________________________
Critical values / negations / corrections: _________________________________
Approved confirmation or staff handoff: ____________________________________
What counts as successful completion: ______________________________________

2. SEPARATE THE CALLER COHORTS
Expected language(s): ______________________________________________________
Expected locales and caller accents: _______________________________________
Single language or within-turn code switching? _____________________________
Consent / lawful audio-use authority / reviewer: ___________________________
English scripts below require separately reviewed language-specific examples.
Do not infer Arabic-English code switching from individual Arabic support.
Do not infer language identification from accent variation.

3. RECORD THE ACTUAL CONFIGURATION
Provider / exact selected model: ___________________________________________
Underlying revision if a receipt proves it, otherwise unknown: _____________
Caller language mode and documented supported set: _________________________
Region / network placement / audio path / encoding: _________________________
Credentials and billing mode ready? ________________________________________
Turn settings / vocabulary hints / SDK version: ____________________________
Burki Azure: BYO key plus same-resource region and explicit supported locale.
No native custom endpoint, deployment selector or automatic language detection.
Azure SDK phrase lists exist, but current Burki editor has no Azure phrase field.
Burki Nova-3: choose an explicit available caller language. The runtime accepts
multi, but no working dashboard entry path was verified here. Confirm a supported
configuration route before using multi. Direct Nova API support is separate;
do not copy a Flux language-hint configuration or assume a custom-code workaround.

4. CONTROLLED FICTIONAL SCRIPTS
A. "Please ask about the Saturday morning pottery session."
B. "Not Saturday. Sunday, if there is a morning session."
C. "My name is Liora Venn. Please read that spelling back to me."
D. Use a synthetic callback number; correct one digit and require readback.
E. Add a reviewed Spanish-English sample only for supported configurations.
Keep identical audio across providers, with the same permitted testing scope.
Do not make paid calls solely to fill this worksheet without separate authority.

5. SAMPLE RESULTS (EMPTY UNTIL EVALUATED)
Sample / cohort: ___________________________________________________________
Expected critical values and correction: ___________________________________
Transcript returned? _______________________________________________________
Observed transcript / request identifier: __________________________________
Wrong name / number / date / negation: ______________________________________
Final structured request correct? __________________________________________
Old value accidentally retained? ___________________________________________
Speech end -> final transcript (TTFS): _____________________________________
Speech end -> first audible agent reply: ___________________________________
Confirmed business action time and receipt: ________________________________
Failure / missing response / retry cost: ___________________________________
Sample count and excluded failures: ________________________________________
Record outcomes per language and accent cohort before combining averages.

6. EXTERNAL BENCHMARK RECORD
Primary URL, immutable revision and read date: ______________________________
Dataset / sample count / language scope: ___________________________________
Exact row identifiers and run dates if stated: _____________________________
Metric definition, judge and confidence intervals: _________________________
Regions, configuration and missing results: ________________________________
The article's Pipecat snapshot uses semantic WER and final-transcript latency.
It is not ordinary exact-word WER, full agent latency or a latest2607 head-to-head.
Do not merge old copied benchmark scores with a rescored current table.

7. COST RECORD
Azure observed public example: Central US, USD, Standard real-time transcription,
October 12, 2026: USD 1 per audio hour, per-second billing. Actual agreement and
region/SKU can differ. Native Azure capabilities do not include every add-on.
Deepgram observed example: streaming PAYG promotional Nova-3 mono USD 0.0048/min,
multi USD 0.0058/min; keyterm prompting separately USD 0.0013/min, October 12.
These are provider rates, not complete Burki call costs or permanent quotes.
Your current dated quote / plan / currency: ________________________________
Billable recognition minutes and add-ons: __________________________________
LLM / synthesis / platform / phone charges: ________________________________
Cost per accepted task, including failures: ________________________________

8. DECISION
Required language contract met? ___________________________________________
Critical-field correctness and correction handling acceptable? _____________
Latency distribution / missing responses acceptable? _______________________
Native configuration and provider acceptance separately verified? __________
Chosen provider / reason / unresolved limits / next review date: ___________
Do not invent a winner from a blank evaluation or broad language coverage.

Primary references:
https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support
https://learn.microsoft.com/en-us/azure/ai-services/speech-service/improve-accuracy-phrase-list
https://developers.deepgram.com/docs/models-languages-overview/
https://azure.microsoft.com/en-us/pricing/details/speech/
https://deepgram.com/pricing
https://github.com/pipecat-ai/stt-benchmark/blob/7e438f94c4b11d9c87c5a03ed7cb2ff819d252d0/README.md