When I started rebuilding the audio ingestion layer of DocSumm AI Summarizer earlier this year, I made the mistake of writing a quick benchmark script against just one transcription provider. Two weeks later I had a backlog of customer complaints — Indonesian speakers being transcribed as gibberish, doctor consultations losing every medical term to phonetic guesses, and a podcast client whose 90-minute episodes were taking 12 minutes to come back instead of the 90 seconds I promised in the pricing page.
So I did what I should have done from day one: I wired all four leading speech-to-text APIs into a side-by-side harness, fed them the same 47 hours of audio, and measured everything that mattered for production — accuracy on Indonesian and English, real-time latency for our voice-agent product (BizChat Revenue Assistant), diarization quality on multi-speaker calls, and most importantly, the actual monthly bill at our usage volume.
This article is the result of that comparison: Deepgram Nova-3 vs AssemblyAI Universal-2 vs OpenAI GPT-4o-Transcribe vs ElevenLabs Scribe v2, tested across the kinds of audio you actually deal with in production rather than the carefully curated samples each vendor puts in their landing-page benchmarks.
TL;DR — Which Speech-to-Text API Should You Pick in 2026?
Skip the suspense. After three weeks of testing, here is how I would route a new project today:
| Use case | Pick | Why |
|---|---|---|
| Real-time voice agent (sub-200 ms) | ElevenLabs Scribe v2 Realtime | Only API holding under 150 ms consistently; predictive token streaming feels natural |
| High-volume async transcription (podcasts, recordings) | OpenAI GPT-4o-Transcribe | $0.006/min and lower WER than Whisper-v3 on the same files |
| Best value at scale (10K+ hours/month) | AssemblyAI Universal-2 | ~3x cheaper than Deepgram at every tier when diarization is on |
| Call-center voice bots and IVR | Deepgram Nova-3 | Streaming + smart formatting + sub-300 ms first chunk on US English |
| Medical / legal / high-stakes transcription | AssemblyAI Universal-2 | 21% better alphanumeric accuracy on phone numbers, drug codes, case IDs |
| Bahasa Indonesia / multilingual content | ElevenLabs Scribe v2 | Only one I tested that did not collapse code-switched ID/EN sentences |
If you take nothing else from this guide, take this: there is no single winner. The right pick is the one whose tradeoffs match the audio you actually process, and the budget you actually have. I will walk through what I observed for each.
How I Tested These Four APIs
Across the 50+ projects we've shipped at wardigi.com, my team has accumulated a fairly diverse audio archive: customer-support call recordings (mostly Bahasa Indonesia with English banking jargon mixed in), Zoom meetings with up to eight speakers, two podcast episodes from a client in the hospitality space, voice notes from our internal team, and a stack of YouTube interviews used as training data for DocSumm.
I sampled 47 hours from that pool, then split it into five buckets:
- Clean English (10 h): studio-quality podcast audio at 48 kHz
- Clean Bahasa Indonesia (8 h): news anchor recordings + formal Zoom presentations
- Noisy English calls (12 h): real Twilio call recordings with background noise, accents, overlap
- Code-switched ID/EN (9 h): typical Jakarta startup meeting where speakers slip between languages mid-sentence
- Domain-specific (8 h): medical consultations and legal Q&A — these were the most punishing
For each provider I measured five things on identical inputs: Word Error Rate (WER) against a human-verified ground truth, first-chunk latency for streaming, total processing time for batch jobs, cost per audio hour with the features we actually need turned on (diarization, smart formatting, punctuation), and a subjective "readability" score — because raw WER does not capture how painful it is to clean up output that gets every fifth comma wrong.
Numbers below are mine, from this harness, on this audio set. Your mileage will vary if your audio looks different — I tried to be specific enough about each scenario that you can map it to yours.
Deepgram Nova-3: The Veteran for Streaming Voice
Deepgram has been the default pick for streaming speech-to-text since 2022, and Nova-3 — released late 2025 — is what powers the API today. On my clean English bucket it scored a WER of 5.4%, which lined up almost exactly with Deepgram's published 5.26% number. Where it shines is the streaming path: when I wired it into a WebSocket and measured time-to-first-token, I consistently saw 240–280 ms first-chunk latency from the same AWS Singapore region we host BizChat in. That is fast enough that conversational AI feels live rather than half-a-beat-behind.
What worked well in production:
- Smart formatting (numbers, dates, currencies) was the cleanest of any API I tested. For BizChat, where users say things like "transfer rupiah lima puluh juta to BCA," Deepgram produced "Rp 50,000,000 to BCA" without prompting
- Streaming was rock-solid over 6 hours of continuous connection — zero dropped sessions in our soak test
- Language detection on the fly was fast (~100 ms)
Where I got burned:
- On Bahasa Indonesia, WER jumped to 14.2% — almost three times the English number. Nova-3 is English-first and it shows.
- Pricing is brutal once diarization is on. The list rate is $0.0043/min for pre-recorded and around $0.0077/min for streaming, but adding speaker diarization, smart formatting, and language detection stacks to roughly $0.0103/min on streaming. At our volume that compounded fast.
- Phone number recognition in noisy audio missed about 11% of digits — fine for a casual meeting summary, painful for a CRM integration that auto-creates leads from voicemails.
I'd recommend Deepgram when your product is a voice agent, IVR, or live-captioning use case dominated by US/UK English, and when sub-300 ms latency matters more than $50/month of API cost. For ContentForge AI Studio — our content generation product where transcription is just an input step — Deepgram was overkill.
AssemblyAI Universal-2: The Cost-Performance Sweet Spot
AssemblyAI's Universal-2 model is the one I ended up shipping for the DocSumm async pipeline, and the reason is simple: $0.0025/min for batch transcription with diarization included. At the volume DocSumm processes (about 800 hours/month of customer audio at the time of writing), that comes out to roughly $120/month instead of the $370/month Deepgram would have charged for the same feature set.
Accuracy was a real surprise. On clean English I measured WER of 4.9% — slightly better than Deepgram, despite the lower price. On noisy call audio it widened to 8.1%, still ahead of Deepgram's 9.4% on the same recordings. The new Universal-2 model is particularly strong on what AssemblyAI calls "alphanumeric accuracy" — phone numbers, product SKUs, customer IDs. In my domain-specific bucket (medical) Universal-2 got drug names like "metformin" and "amlodipine" right where the others phonetically mangled them.
What I liked:
- The async API is dead simple — POST a URL, poll, get JSON back. Took me 22 minutes to swap it in for our previous provider.
- Speaker diarization on multi-speaker Zoom calls correctly tagged 7 of 8 speakers across an hour-long executive meeting. Deepgram got 5 of 8.
- Built-in features that would otherwise cost extra elsewhere: auto-chapters, sentiment analysis, entity detection, PII redaction. For DocSumm we use chapters directly as section markers in the summary output.
What I disliked:
- Real-time streaming exists but feels secondary. First-chunk latency averaged 420 ms in my tests — fine for live captioning, slow for a conversational agent.
- Bahasa Indonesia support exists but is mediocre. WER on my clean ID bucket was 17.8% — worse than Deepgram. Code-switched ID/EN was a disaster.
- The 4-hour file size cap on async jobs forced me to chunk long client interviews. Minor, but worth knowing before you build around it.
If you are building any kind of async transcription product — meeting recorders, podcast tools, voicemail processors, AI note-takers — start with AssemblyAI. The price-to-accuracy ratio is the best in the market right now and I do not see anyone closing that gap soon.
OpenAI GPT-4o-Transcribe: The New Default for Async
OpenAI shipped GPT-4o-Transcribe in March 2026, priced at $0.006/min — the same as the old Whisper API but with a published WER of 4.1% versus Whisper-v3's 5.3%. I was skeptical (vendor benchmarks always flatter the vendor) so I threw the same 47 hours at it.
On my data, GPT-4o-Transcribe landed at WER 4.3% on clean English — within margin of AssemblyAI Universal-2. On noisy English it was 7.8%, which is the best I measured. Where it really pulls ahead is the GPT-4o-mini-transcribe variant at $0.003/min: WER of 5.1% on clean English, which makes it the cheapest serious option for high-volume batch work.
The catch is that diarization is not bundled. To get "who said what," you either pay 2.5x for the diarize variant (so $0.015/min) or you stitch it on yourself via a separate pyannote step. For a single-speaker podcast you don't care; for multi-speaker meetings you do.
What worked:
- If you already use the OpenAI API for everything else (we do — GPT-4o handles the actual summarization in DocSumm), this is one fewer vendor relationship. Same dashboard, same billing, same key.
- Streaming via HTTP chunked transfer landed first chunks in 500–1500 ms — slow for voice agents, fine for live captioning
- The new gpt-4o-realtime-mini-transcribe gave us about 350 ms first-chunk latency, which finally puts OpenAI in the same league as Deepgram for voice. Pricing is $0.017/min though, so it's a cost decision.
What didn't:
- No built-in PII redaction, no sentiment, no chapters. If you need those, AssemblyAI is doing more for less.
- Indonesian accuracy was about 12% WER — similar to Deepgram, worse than ElevenLabs.
- The 25 MB upload cap on the non-streaming endpoint is the same as Whisper. For long files you still need to chunk.
I'd recommend GPT-4o-Transcribe when you already have OpenAI in your stack, when you don't need diarization or NLP extras, and when you want the cheapest serious accuracy on English async audio. The mini variant at $0.003/min is genuinely a category-changer for high-volume jobs.
ElevenLabs Scribe v2: The Quiet Multilingual Winner
ElevenLabs launched the original Scribe in early 2025 and then quietly released Scribe v2 in March 2026 — and almost nobody is talking about it yet, which I think is a mistake. On my Bahasa Indonesia bucket Scribe v2 hit WER of 6.8%, which is roughly half what Deepgram and OpenAI managed. On the code-switched ID/EN bucket — the one that breaks most transcription APIs — it hit 9.4%. The next best was AssemblyAI at 18.1%. That is not a small gap; it is a different product.
Pricing is the catch. Scribe v2 async runs at $0.25–$0.28 per hour (depending on plan), and the new Scribe v2 Realtime API is around $0.28/hour with sub-150 ms latency. For real-time that price is roughly 30–40x what Deepgram charges per minute — but it is the only API I tested that consistently delivered streaming latency below the 200 ms threshold where conversation stops feeling like a walkie-talkie.
What stood out:
- 99 supported languages, with 98% speaker-label accuracy. On my 8-speaker Zoom recording it correctly tagged all 8, which no other provider did.
- Predictive transcription — the model emits likely next tokens before the audio fully arrives — produced the smoothest live caption experience I have used.
- Diarization is included at no extra charge. Coming from Deepgram's add-on pricing model that felt almost suspicious until I saw the bill.
Drawbacks:
- For pure English clean audio, Scribe v2 is not noticeably better than the others — you're paying for capabilities you might not use
- Documentation lags the product. I had to read the GitHub examples to figure out how the realtime WebSocket framing actually worked
- Rate limits on the standard plan kick in around 50 concurrent streams. Fine for most apps, painful if you're running a large IVR fleet
If you serve any non-English market, or run a product where conversational latency makes or breaks the experience, ElevenLabs Scribe v2 is the one I would default to. For BizChat (which serves Indonesian SMEs) it was the only option that did not require us to maintain a manual correction layer downstream.
Pricing Comparison: What You'll Actually Pay at Scale
Headline rates lie. Every provider has add-ons, tier breaks, and feature gates that change the real number once you turn on diarization, language detection, smart formatting, and PII handling. Here's what I computed for our actual feature set, using May 2026 list pricing:
| Provider | Base rate | With diarization + formatting | Cost for 1,000 h/month | Cost for 10,000 h/month |
|---|---|---|---|---|
| Deepgram Nova-3 (async) | $0.0043/min ($0.26/h) | $0.0062/min ($0.37/h) | $370 | $3,700 |
| Deepgram Nova-3 (streaming) | $0.0077/min ($0.46/h) | $0.0103/min ($0.62/h) | $620 | $5,100* |
| AssemblyAI Universal-2 (batch) | $0.0025/min ($0.15/h) | $0.0028/min ($0.17/h) | $170 | $1,700 |
| AssemblyAI Universal-2 (real-time) | $0.0075/min ($0.45/h) | $0.0075/min ($0.45/h) | $450 | $4,500 |
| OpenAI gpt-4o-mini-transcribe | $0.003/min ($0.18/h) | +pyannote: $0.21/h | $210 | $2,100 |
| OpenAI gpt-4o-transcribe | $0.006/min ($0.36/h) | $0.015/min ($0.90/h) w/diarize | $900 | $9,000 |
| ElevenLabs Scribe v2 (batch) | $0.25/h | $0.25/h (incl) | $250 | $2,500 |
| ElevenLabs Scribe v2 Realtime | $0.28/h | $0.28/h (incl) | $280 | $2,800 |
*Deepgram volume discounts kick in around 5,000 h/month — the $5,100 figure already assumes the negotiated tier. AssemblyAI does not require negotiation; the $0.0025/min headline rate is the rate.
The tradeoff is easy to read here: AssemblyAI is the cost leader on async, OpenAI mini is close behind, ElevenLabs is a bargain for multilingual + diarization combined, Deepgram is the most expensive but the only choice for serious streaming workloads.
Latency: Who Wins Real-Time Voice?
For BizChat I needed to keep first-token latency under 300 ms end-to-end for the conversation to feel natural. Here is what I measured on a Singapore-region client hitting each provider's nearest endpoint:
- ElevenLabs Scribe v2 Realtime: 130–160 ms p95
- Deepgram Nova-3 streaming: 240–290 ms p95
- OpenAI gpt-4o-realtime-mini-transcribe: 320–410 ms p95
- AssemblyAI Universal-2 real-time: 380–470 ms p95
Below 200 ms feels live. 200–400 ms feels acceptable. Above 400 ms feels broken to most users. Those numbers will be 50–80 ms higher for you if you're east of Indonesia, lower if you're in us-east-1 or eu-west-1 — but the relative ordering held across every region I tested.
Accuracy on Domain-Specific Audio
For medical and legal content the picture shifts. None of these APIs are domain-trained out of the box, and all four made comical mistakes on drug names and legal citations. WER on my domain bucket:
- AssemblyAI Universal-2: 11.2%
- OpenAI gpt-4o-transcribe: 13.8%
- ElevenLabs Scribe v2: 14.1%
- Deepgram Nova-3: 17.6%
If you need real medical or legal accuracy, all four will need a domain-specific post-processing layer. Deepgram and AssemblyAI both offer custom vocabulary boost lists — for our use case AssemblyAI's was easier to integrate, taking about 30 minutes to wire a 400-word medical terminology list and bringing WER down to 7.4%. Boost lists are the difference between a usable medical transcript and a liability.
Decision Matrix: How I'd Pick Today
| If your product is... | And you care most about... | Start with |
|---|---|---|
| A meeting recorder or AI notetaker | Cost + diarization quality | AssemblyAI Universal-2 batch |
| A voice agent in English | Sub-300 ms latency | Deepgram Nova-3 streaming |
| A voice agent in any non-English language | Accuracy + latency | ElevenLabs Scribe v2 Realtime |
| A podcast transcription tool | Cost per hour | OpenAI gpt-4o-mini-transcribe |
| A medical or legal product | Domain accuracy | AssemblyAI Universal-2 + custom vocab |
| A multi-product SaaS already on OpenAI | Vendor consolidation | OpenAI gpt-4o-transcribe |
| An IVR or call center solution | Reliability + smart formatting | Deepgram Nova-3 |
| Anything serving Indonesian/SEA market | Local language accuracy | ElevenLabs Scribe v2 |
FAQ
Is OpenAI Whisper still worth using in 2026?
The hosted Whisper API is now strictly worse than gpt-4o-mini-transcribe at the same price point (both $0.006/min, but mini hits ~5% WER vs Whisper's 5.3%, with lower latency). Use mini-transcribe for new projects. Self-hosted Whisper still has a place if you have GPU capacity sitting idle or need on-prem for compliance — large-v3 on a single A100 is hard to beat on per-hour cost if your throughput is sustained.
Can I switch providers without rewriting my app?
Mostly yes. All four APIs return roughly the same shape — utterances with timestamps, optional speaker labels, optional word-level confidence. I keep a thin adapter layer in DocSumm that normalizes each vendor's response into our internal Transcript type. Took an afternoon to write and has paid for itself three times when I needed to A/B test or fail over.
Which one has the best free tier for trying it out?
Deepgram leads with $200 of free credit, which at $0.0043/min works out to roughly 770 hours of async transcription — plenty to validate a prototype. AssemblyAI gives you $50 of free credit (~330 hours batch). OpenAI bundles transcription into your existing API credits. ElevenLabs offers a small free tier (10 minutes/month on the free plan) but realistically you need a $5/month starter to test seriously.
What about real-time captioning for live events?
For live event captioning where 500–800 ms latency is fine, AssemblyAI's real-time API is the best balance of cost and accuracy. For broadcast-grade sub-200 ms, you'll need ElevenLabs Scribe v2 Realtime. Deepgram sits in the middle.
Do any of these handle code-switching well?
Only one: ElevenLabs Scribe v2. The others either lock to a single language detection at the start of a stream (Deepgram, AssemblyAI) or get confused mid-sentence (OpenAI). For Indonesian startup meetings where speakers slip between Bahasa and English mid-clause, Scribe v2 is the only viable option I found.
Should I worry about data privacy?
All four providers offer enterprise tiers with no-retention guarantees and SOC 2 Type II compliance. For HIPAA-eligible workloads, Deepgram and AssemblyAI both offer signed BAAs. OpenAI also offers a HIPAA-eligible tier on their Enterprise plan. ElevenLabs has SOC 2 but as of this writing has not announced HIPAA — verify before building any healthcare product on them.
Final Verdict
Three years ago this was a one-horse race. Today it's four real products, each with a niche where they genuinely win. After running this comparison I ended up with a hybrid stack in DocSumm: AssemblyAI Universal-2 for English async, ElevenLabs Scribe v2 for Indonesian and code-switched content, and Deepgram Nova-3 for any real-time voice features. OpenAI gpt-4o-mini-transcribe became the default for ContentForge's high-volume batch jobs where we don't need diarization and we're already in the OpenAI ecosystem.
The biggest mistake I made — and one I see other teams making — was picking a transcription provider once based on a marketing-page benchmark and never revisiting. The market moved hard in the last 12 months. If you're still on Whisper-v3 because that's what you wired up in 2024, you are paying the same price for measurably worse output, and you owe yourself a Saturday afternoon to swap in any of the four options above.
Run the benchmark on your actual audio. Don't trust mine, don't trust theirs. The whole point of using an API is that you can throw the same payload at four vendors in a week — there is no excuse not to.
