If you're building transcription into an app, a call-center tool, a note-taker or a voice agent, you need a speech to text API, not a dictation app. This post compares six of the most-used options on what their own pages say: models, streaming support, notable features and list price, all as of October 2026.
First, a plain disclosure, because people find this post from our site: LumeVoice is not an API. It's a dictation app for people on Mac, Windows, iPhone and Android. You hold a key, talk, and clean text appears in whatever app you're using. There's no developer API, SDK or per-minute plan. If you're a developer who just wants to talk to your editor or coding agent instead of typing, that's what our developers page is about. If you're building speech into software, read on.
We haven't benchmarked these APIs, and we don't quote accuracy scores. Published results depend on the test set, and the only score that matters is the one on your audio.
At a glance#
| API | Models (Oct 2026) | Streaming | List price | Free to start |
|---|---|---|---|---|
| OpenAI | gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, whisper-1 | Yes (realtime models) | $0.003–$0.006/min for files; $0.017/min realtime | Open-source Whisper to self-host |
| Deepgram | Nova-3 (mono and multilingual), Flux | Yes | $0.0043–$0.0052/min pre-recorded; $0.0048–$0.0078/min streaming | $200 credit |
| AssemblyAI | Universal-3.5 Pro, Universal-2, Universal-Streaming | Yes | $0.15–$0.21/hr files; from $0.15/hr streaming | $50 credit |
| Google Cloud | Speech-to-Text V2 | Yes | $0.016/min standard (first tier); $0.003/min dynamic batch | See free tier note |
| Azure AI Speech | Real-time, Fast and Batch transcription, MAI Transcribe | Yes | Per audio hour, see calculator | 5 hrs/month (free tier) |
| Amazon Transcribe | Batch and streaming | Yes | Per minute by region | 60 min/month for 12 months |
Swipe to see all columns
What to decide before comparing prices#
Files or live audio? Batch (pre-recorded) transcription is cheaper and simpler. Streaming is for captions, voice agents and anything a person waits on. Some models only do one.
Which languages? Many vendors have separate English and multilingual models at different prices. Check the language list for the exact model.
Which extras? Speaker labels (diarization), redaction of personal information, custom vocabulary or key terms, timestamps and translation are priced and supported differently everywhere.
Where does the data go? Read each vendor's data-use and retention terms, and whether audio may be used to improve models by default. Regulated industries should check for agreements such as a BAA before sending any audio.
How is it billed? Per minute, per hour, per second or per token; with or without minimums; streaming billed on audio or on connection time. These details move your real cost more than the headline rate.
1. OpenAI: simple file transcription, plus realtime#
OpenAI speech to text
Verdict: Easiest Start, Especially if You Already Use OpenAI
OpenAI's pricing page lists several transcription models with an estimated cost per minute: gpt-4o-mini-transcribe about $0.003, gpt-transcribe about $0.0045, gpt-4o-transcribe and whisper-1 about $0.006, and gpt-4o-transcribe-diarize (with speaker labels) about $0.006. For live audio, gpt-live-transcribe and gpt-realtime-whisper are about $0.017 a minute, and gpt-realtime-translate about $0.034 (OpenAI pricing).
- Good at: one API key for transcription and the rest of your AI stack; a diarization model; the Whisper family is also open source, so you can self-host (Whisper on GitHub).
- Watch for: the newer models are priced in tokens, so per-minute figures are OpenAI's estimates.
2. Deepgram: built around streaming#
Deepgram
Verdict: Strong Choice for Live Voice Apps
Deepgram's pay-as-you-go page lists Nova-3 Monolingual at $0.0043 a minute pre-recorded and $0.0048 streaming, Nova-3 Multilingual at $0.0052 and $0.0058, and its streaming-only Flux models at $0.0065 (English) and $0.0078 (multilingual) a minute. Some streaming rates were shown as limited-time promotional prices, with the standard rate crossed out. New accounts get $200 in credit (Deepgram pricing).
- Good at: real-time transcription for voice agents and live products; separate mono and multilingual models.
- Watch for: promotional streaming rates may change; check the standard rate for long-term budgeting.
3. AssemblyAI: priced per hour, feature-rich#
AssemblyAI
Verdict: Best for Files With Lots of Extras
AssemblyAI lists Universal-3.5 Pro at $0.21 an hour and Universal-2 at $0.15 an hour for pre-recorded audio. Streaming is $0.15 an hour for Universal-Streaming (English or multilingual) and from $0.45 an hour for Universal-3.6 Pro Realtime; streaming is billed on connection time, not audio sent. New accounts get $50 of credit with no card (AssemblyAI pricing).
- Good at: extras on the same platform: speaker diarization, keyterms prompting, entity detection, PII redaction (47+ languages on Universal-2), translation to 100+ languages for pre-recorded audio, and a Medical Mode for English, Spanish, German and French.
- Watch for: connection-time billing on streaming; close idle sessions.
4. Google Cloud Speech-to-Text#
Google Cloud Speech-to-Text
Verdict: Best Fit if You Already Run on Google Cloud
Google's pricing page lists V2 standard recognition at $0.016 a minute for the first 500,000 minutes a month, falling in tiers to $0.004 at 2,000,000 minutes and above. Dynamic batch, for audio you can wait for (Google says within 24 hours), is listed at $0.003 a minute (Google Cloud pricing). Google's free-tools page lists 60 minutes of audio a month free; confirm how it applies to the API version you use.
- Good at: volume discounts, a cheap batch path, and keeping billing and IAM inside Google Cloud.
- Watch for: pricing differs between API versions and models, so price the exact one you'll call.
5. Azure AI Speech#
Azure AI Speech
Verdict: Best Fit for Microsoft Shops
Azure's speech pricing page lists Real-time, Fast and Batch transcription, plus MAI Transcribe models, each billed per audio hour in one-second increments. The free (F0) tier includes 5 audio hours a month of standard real-time transcription, shared with Custom speech; batch isn't on the free tier (Azure Speech pricing). Azure shows rates by region, so use its pricing calculator for a quote.
- Good at: custom speech models, enterprise agreements, and data residency in Azure regions.
- Watch for: rates vary by region and commitment tier.
6. Amazon Transcribe#
Amazon Transcribe
Verdict: Best Fit if Your Audio Already Lives in AWS
Amazon Transcribe offers batch and streaming transcription, billed per second with no minimum, with rates set per region and in volume tiers. The free tier gives 60 minutes a month for 12 months from your first request; it doesn't cover automatic content redaction or custom language models (Amazon Transcribe pricing).
- Good at: S3-based pipelines, custom language models and redaction, inside AWS's security and billing.
- Watch for: pick your region on the pricing page to see the actual rate.
Common mistakes when choosing#
Comparing headline rates only. A cheaper per-minute rate can cost more once you add diarization, redaction or a multilingual model, or when streaming bills for idle connection time.
Testing on clean audio. Vendor demos use clear speech. Your users have speakerphones, accents, cross-talk and jargon. Test with that.
Ignoring promotional pricing. Deepgram showed some streaming rates as limited-time offers in October 2026. Budget on the standard rate if you're planning a year ahead.
Forgetting the free tier ends. AWS's free minutes last 12 months; credit-based offers run out with use.
Locking in too early. Keep your transcription call behind a small interface in your code so you can switch vendors if prices or quality change. The market moves quickly, and models released this year may be retired next year.
How to pick in an afternoon#
- Collect 20 real clips: your accents, your noise, your jargon, both short and long.
- Shortlist three based on the decisions above (live vs files, languages, extras, cloud).
- Run the same clips through each using free credit. Read the transcripts side by side, paying attention to names, numbers and punctuation.
- Price your actual month, including extras, idle streaming time and volume tiers.
- Read the data terms before sending any customer audio.
Where LumeVoice fits (and doesn't)#
It doesn't, for building software. LumeVoice runs Whisper large-v3-turbo on Groq, with Mistral's Voxtral as a fallback, behind a dictation app, and none of that is exposed as an API. If what you actually want is to dictate faster, into Slack, email, your IDE or a coding agent, that's where it helps.
Want to dictate, not build?
LumeVoice is a dictation app, not an API: hold a key, talk, and clean text lands in your editor, terminal or chat. Every new account gets 3,000 free words, no card required.
