Direct answer: for US regions, standard models, transcription minutes only, excluding storage/LLM/egress, 2026 monthly speech-to-text costs at 100k minutes are roughly: Deepgram Nova-3 770 (Growth 650), Azure Speech S1 1,666.67 (2,000-hour commitment 1,600), Google STT 1,600 (1-year commitment 1,440); at 10k minutes: Deepgram 77, Azure 166.67, Google 160; at 1M minutes: Deepgram 7,700 (Growth 6,500), Azure 10,833 (10,000-hour commitment), Google 13,000 (1-year 11,700). Deepgram has the lowest per-second rate; Azure’s $1/hour plus commitment tiers fit steady volume; Google’s “data logging” choice moves the rate from 0.016 to 0.024/minute, a 50% difference. Confirm data logging, diarization, concurrency and commitments before comparing per-minute rates.
This guide is for engineering, data and procurement teams choosing a speech-to-text API or putting call recordings/meetings/subtitles/QA into a pipeline. Data cut-off: 5 November 2026; prices are the three vendors’ official US list prices/public price APIs, in USD, excluding tax; the TCO is an editorial calculation, not a vendor quote. All three bill per second/minute of audio, but the real differences are standard/enhanced models, data logging, diarization, commitment tiers and minimums.
Deepgram Nova-3 monolingual is 0.0077/minute (Growth 0.0065), billed per second with no minimum; Azure S1 is 1.00/hour, 2,000-hour commitment 1,600/month (about 0.80/hour), 10,000-hour 6,500/month (about 0.65/hour); Google STT standard is 0.016/minute with data logging or 0.024 without, with 60 minutes/month free. At 100k minutes Deepgram 770 is lowest, Azure commitment 1,600, Google 1,600; at 1M minutes Deepgram Growth 6,500 is lowest.
Three models: per second, per hour + commitment, per minute + data logging
Deepgram: Nova-3 monolingual 0.0077/minute (PAYG), Growth 0.0065; multilingual 0.0092/0.0078; Flux STT English 0.0077/0.0065; per-second billing with no block rounding; $200 trial credit (about 43k minutes); Growth starts at 4,000/year with up to about 20% off; diarization 0.0020/minute, redaction 0.0020, keyterm 0.0013, entity 0.0017. Azure: Speech to Text S1 1.00/hour (per second), 5 free hours/month; 2,000-hour commitment 1,600/month, overage 0.80/hour; 10,000-hour 6,500/month, overage 0.65/hour; enhanced features 0.30/hour; custom model hosting 0.05375/hour. Google: STT standard with data logging 0.016/minute (first 500k), without data logging 0.024/minute; 500k–1M 0.010; 1M–2M 0.008; 2M+ 0.004; 60 minutes/month free; 1-year commitment 0.0144/0.009/0.0072/0.0036; 3-year lower.
| Platform | Billing unit | Standard rate | Commitment/discount | Free/trial | Main boundary |
|---|---|---|---|---|---|
| Deepgram | Second (shown per minute) | Nova-3 monolingual 0.0077/minute; multilingual 0.0092 | Growth 0.0065/0.0078; from 4,000/year | $200 credit (about 43k minutes) | Per second, no rounding; add-ons separate |
| Azure Speech | Second (shown per hour) | S1 1.00/hour | 2K hours 1,600/month; 10K hours 6,500/month | 5 hours/month | Commitment/enhanced features separate |
| Google STT | Second (shown per minute) | With logging 0.016/minute; without 0.024 | 1-year 0.0144/0.009; 3-year lower | 60 minutes/month | Data-logging choice changes 50% |
Sources: Deepgram Pricing, Azure Speech Pricing and Google Cloud Speech-to-Text Pricing (all accessed 5 November 2026). This is editorial grouping; “standard rate” covers only public standard models, excluding enhanced/custom/batch/Voice Agent; actual prices vary by region, currency, commitment and add-ons.
The 2026 official prices: per minute, per hour
Deepgram Nova-3 monolingual 0.0077/minute = 0.462/hour, Growth 0.0065 = 0.39/hour; multilingual 0.0092 = 0.552/hour. Azure S1 1.00/hour, 2K commitment 1,600/2,000h = 0.80/hour, 10K commitment 6,500/10,000h = 0.65/hour; enhanced features 0.30/hour. Google with logging 0.016/minute = 0.96/hour, 1-year 0.0144 = 0.864/hour, 3-year 0.0128 = 0.768/hour; without logging 0.024/minute = 1.44/hour, 1-year 0.0216 = 1.296/hour, 3-year 0.0192 = 1.152/hour.
| Basis | Deepgram | Azure Speech | Google STT |
|---|---|---|---|
| Entry/minute | 0.0077 (monolingual) | 0.01667 (1.00/hour) | 0.016 (with logging)/0.024 (without) |
| Entry/hour | 0.462 | 1.00 | 0.96 (with logging)/1.44 (without) |
| Commitment/discount | Growth 0.0065/minute = 0.39/hour | 2K 0.80/hour; 10K 0.65/hour | 1-year 0.864/hour; 3-year 0.768/hour (with logging) |
| Free allowance | $200 credit (about 43k minutes) | 5 hours/month | 60 minutes/month |
| Billing granularity | Per second, no rounding | Per second | Per second |
Sources: Deepgram Pricing (0.0077/0.0065, multilingual 0.0092/0.0078, $200 credit), Azure Speech Pricing (S1 1.00/hour, 2K 1,600, 10K 6,500, 5 free hours, enhanced 0.30/hour) and Google STT Pricing (with/without logging 0.016/0.024, 1-year 0.0144/0.0216, 3-year 0.0128/0.0192, 60 free minutes). Accessed 5 November 2026; USD, excluding tax; hourly rates are editorial conversions.
10k / 100k / 1M minutes: monthly TCO
For standard models, transcription only: 10k minutes, Deepgram 77, Deepgram Growth 65, Azure S1 166.67, Google 160 (1-year 144); 100k minutes, Deepgram 770/650, Azure 1,666.67 (2K commitment 1,600), Google 1,600 (1-year 1,440); 1M minutes, Deepgram 7,700/6,500, Azure 10,833.33 (10K commitment), Google 13,000 (1-year 11,700). Deepgram is lowest in all three bands; Azure’s commitment approaches/beats Google around 100k minutes; Google standard is close to Azure S1, while the no-logging version is 50% more.
| Monthly minutes | Deepgram PAYG | Deepgram Growth | Azure S1 | Azure commitment | Google standard | Google 1-year |
|---|---|---|---|---|---|---|
| 10k | 77 | 65 | 166.67 | 166.67 (below 2K) | 160 | 144 |
| 100k | 770 | 650 | 1,666.67 | 1,600 (2K) | 1,600 | 1,440 |
| 1M | 7,700 | 6,500 | 16,666.67 | 10,833.33 (10K) | 13,000 | 11,700 |
Editorial calculation, not a quote. Deepgram = minutes×0.0077/0.0065; Azure S1 = minutes/60×1.00; Azure 2K = 1,600 (≤2,000 hours), 10K = 6,500 + hours over 10,000×0.65; Google with logging standard = first 500k×0.016 + 500k×0.010 (1M), 1-year = 0.0144/0.009; excludes tax, storage, LLM/NLP, egress, enhanced features, custom models and minimums. Deepgram Growth’s 4,000/year cash commitment is not amortized into the monthly figure.
Enhanced features: diarization, redaction, keyterms, logging
The common add-ons are diarization/redaction/keyterms. Deepgram: diarization 0.0020/minute, redaction 0.0020, keyterm 0.0013, entity 0.0017; at 100k minutes diarization adds 200/month. Azure: enhanced features 0.30/hour (about 0.005/minute), 100k minutes +500, covering language ID/diarization/pronunciation; custom model hosting 0.05375/hour. Google: data-logging choice directly changes the rate (0.016 vs 0.024); whether diarization is included/extra depends on the model/feature page and is not claimed as free here. Storage, LLM summarization, translation and egress are separate.
| Item | Deepgram | Azure | |
|---|---|---|---|
| Diarization | 0.0020/minute | Enhanced 0.30/hour (about 0.005/minute) | Confirm on feature/model page |
| Redaction/keyterms | Redaction 0.0020; keyterm 0.0013; entity 0.0017 | Enhanced included/separate | Confirm on feature page |
| Data logging | Not applicable | Not applicable | With logging 0.016/minute; without 0.024 |
| Custom model | Enterprise custom | Hosting 0.05375/hour; training separate | Confirm on model/training page |
| Storage/LLM/egress | Separate | Separate | Separate |
Sources: Deepgram Pricing (add-ons 0.0020/0.0013/0.0017), Azure Speech Pricing (enhanced 0.30/hour, custom hosting 0.05375/hour) and Google STT Pricing (with/without logging rates). Accessed 5 November 2026; USD, excluding tax; “confirm on feature page” means no single public price is listed and is not guessed here.
Which option fits which scenario?
Low volume/price-first/per-second: Deepgram; steady volume/Azure ecosystem/commitment/enhanced features: Azure; already on GCP and can accept data logging with 1/3-year commitment: Google; cannot log data: Google no-logging 0.024/minute, so check whether it still beats Azure/Deepgram. Heavy diarization/redaction: add Deepgram 0.002/0.002 or Azure 0.30/hour; verify output formats, concurrency and region in the POC.
| Scenario | First choice | Why | Watch |
|---|---|---|---|
| Low volume/price-first | Deepgram | 0.0077/minute, per second no rounding | Add-ons separate |
| Steady volume/commitment | Azure | 2K/10K hours at 0.80/0.65/hour | Commitment/overage/enhanced separate |
| GCP ecosystem/1–3 years | 1-year from 0.0144/minute | Data logging changes 50% | |
| Cannot log data | Google no-logging/others | No-logging 0.024/minute | 50% above with-logging |
| Diarization/redaction | Compare add-ons | Deepgram 0.002/minute; Azure 0.30/hour | Calculate the real feature mix |
| Custom model | Azure/Enterprise | Hosting/training separate | Training/hosting/concurrency |
Sources: Deepgram/Azure/Google official pricing pages (accessed 5 November 2026). This is an editorial decision matrix, not vendor advice; actual accuracy, latency, language coverage, diarization quality, concurrency and support response were not tested.
What should procurement and acceptance specify?
At least eight things: monthly audio hours/peak, real-time/batch, language/accent, diarization/redaction/keyterms, data logging/retention, concurrency/latency, output format/timestamps and commitment/minimum. Use a 2–4 week POC with real audio and record WER/keyterm accuracy, P95 latency, concurrency, diarization, failures/retries and the bill; do not compare only “per minute”.
- Audio volume: monthly hours/peak, real-time/batch, average duration; billed per second.
- Language/accent: monolingual/multilingual, dialect, domain terms; model prices differ.
- Features: diarization, redaction, keyterms, timestamps, punctuation/formatting.
- Data logging: whether logging is allowed; Google’s with/without logging differs by 50%.
- Concurrency/latency: real-time concurrency, P95, queueing/limits; commitment/Enterprise differ.
- Output: JSON/SRT/VTT, timestamp precision, batch export.
- Commitment: Deepgram Growth 4,000/year, Azure 2K/10K hours, Google 1/3-year.
- POC: 2–4 weeks of real audio; record WER, P95, concurrency, diarization, failures and the bill.
FAQ
100k minutes: what is cheapest?
Deepgram Growth 650, Deepgram PAYG 770, Azure 2K commitment 1,600, Google 1-year 1,440, Google standard 1,600, Azure S1 1,666.67. Deepgram is lowest; Azure/Google commitment terms depend on the minimum.
What about 1M minutes?
Deepgram Growth 6,500, Azure 10K commitment 10,833.33, Google 1-year 11,700, Google standard 13,000, Azure S1 16,666.67. Deepgram is lowest; Azure’s commitment is below Google’s 1-year.
Why does Google have two prices?
With data logging 0.016/minute, without 0.024/minute, a 50% difference; 1/3-year commitments reduce further. If logging is not allowed, calculate at 0.024 and compare with Deepgram/Azure.
How does Azure’s commitment work?
2,000 hours 1,600/month (about 0.80/hour), 10,000 hours 6,500/month (about 0.65/hour); overage at 0.80/0.65 per hour. 100k minutes is about 1,666.7 hours, so the 2K commitment is slightly below S1 1,666.67.
Is Deepgram Growth always cheaper?
The unit rate is 0.0065 vs 0.0077, but Growth is a 4,000/year prepaid commitment; 100k minutes costs 650, far below the annual commitment. Low volume should use PAYG; negotiate Growth/Enterprise only at sufficient annual volume.
How much does diarization add?
Deepgram 0.0020/minute, 100k minutes +200; Azure enhanced 0.30/hour, 100k minutes +500; Google depends on the feature/model page. Add it for meetings/QA.
Are real-time and batch the same price?
This article uses standard transcription; real-time/batch/fast transcription may have different SKUs (Azure has a Fast Transcription promo, Google has different SKUs). Confirm the actual API/model.
What other hidden costs exist?
Storage, LLM summarization/translation/NLP, egress, concurrency/limits, custom model training/hosting, commitment minimums, tax and support. This article counts standard transcription minutes only.
Research method and limitations
This article separates “official facts” from “editorial modelling”: Deepgram/Azure/Google standard transcription rates, commitments, free allowances and add-on prices come from the three official pricing pages/public price APIs (US); the 10k/100k/1M-minute TCO is arithmetic under standard models and transcription minutes only, not a vendor quote. It does not compare actual accuracy, latency, language coverage, diarization quality, concurrency or support response.
Eight limitations: (1) only US public list prices are used, excluding tax, region/currency, private pricing and discounts; (2) only standard transcription is counted, excluding real-time/batch/Fast Transcription SKUs; (3) storage, LLM/NLP, translation, egress, concurrency/limits and custom models are excluded; (4) only 10k/100k/1M minutes are modelled, while actual volume/peak changes the commitment tier; (5) Deepgram Growth 4,000/year, Azure 2K/10K hours and Google 1/3-year commitments are not amortized; (6) accuracy, latency, language coverage and diarization quality were not tested; (7) data logging/retention/compliance differences are not costed; (8) AssemblyAI/OpenAI/Whisper/other STT were not compared. Treat this as a unit-alignment template and replace assumptions with a 2–4 week real-audio POC.
Sources and basis
- Deepgram, Pricing: deepgram.com/pricing. Nova-3 monolingual 0.0077/minute, Growth 0.0065; multilingual 0.0092/0.0078; $200 credit; Growth 4,000/year; diarization 0.0020, redaction 0.0020, keyterm 0.0013, entity 0.0017. Accessed 5 November 2026.
- Microsoft Azure, Speech Pricing: azure.microsoft.com/.../speech-services. S1 1.00/hour, 5 free hours/month; 2K commitment 1,600/month, 10K 6,500/month; enhanced 0.30/hour; custom hosting 0.05375/hour. Accessed 5 November 2026.
- Google Cloud, Speech-to-Text Pricing: cloud.google.com/speech-to-text/pricing. With logging 0.016/minute, without 0.024; 500k–1M 0.010; 1M–2M 0.008; 2M+ 0.004; 60 minutes/month free; 1-year 0.0144/0.009/0.0072/0.0036; 3-year lower. Accessed 5 November 2026.
Data cut-off 5 November 2026; prices are public list prices excluding tax; all models are editorial calculations. This is not procurement advice.