By VexaScribe Editorial · Pricing verified July 8, 2026 · Benchmark data added
Best Transcription APIs for Developers in 2026 (12 Tested)
If you're building speech-to-text into your product, the API landscape has consolidated in 2026. OpenAI's Whisper commoditized multilingual transcription, but purpose-built engines from Deepgram, AssemblyAI, and Speechmatics now beat Whisper on English accuracy, latency, and diarization. We benchmarked 12 APIs on English WER, accented speech, noisy audio, streaming latency, pricing, and SDK ergonomics so you can pick the right one without three weeks of trial integrations.
To back this up with independent numbers, we ran our own July 2026 first-hand benchmark of 14 speech-to-text models across 16 standard datasets and 904 audio files. Speechmatics Melia-1 led aggregate WER at 6.4% (best of 14 tested, $0.24/hr); AssemblyAI Universal-3.5 Pro led promptable AI-transcription APIs at 7.0% WER; OpenAI GPT-4o Transcribe collapsed to 43.8% WER on long-form financial earnings calls; Deepgram Nova-3 tied for best on English meeting audio and remains the right choice for real-time streaming. Full per-provider results below.
The short answer: Deepgram Nova-3 for production English workloads, AssemblyAI for the cleanest developer experience, OpenAI's Whisper API when you need 99 languages and can live with batch, self-hosted faster-whisper when you need data control or 100× real-time for pennies.
Quick Decision Rule:
- • Real-time English product → Deepgram Nova-3 ($0.0077/min streaming)
- • Rich audio intelligence (summaries, sentiment, PII) → AssemblyAI
- • 99 languages, batch-tolerant → OpenAI Whisper API
- • EU data residency → Gladia or Speechmatics
- • AWS-native call analytics → Amazon Transcribe Call Analytics
- • Cheapest hosted Whisper → Groq (~$0.02/hr)
- • Full data control / offline → faster-whisper on your GPU
Disclosure: VexaScribe does not currently offer a public transcription API — this comparison is written for developers choosing between third-party APIs. We have no commercial incentive to favor any provider below. Pricing was verified on official pricing pages on July 8, 2026; rates change frequently. Benchmark numbers combine public WER reports, OpenSLR/LibriSpeech evaluations, and our own spot-checks on 30 minutes of mixed-domain audio. The Novascribe 2026 Benchmark section below reports our own independent July 2026 test: 904 audio files across 16 standard benchmarks, run through the official API of every provider covered, WER computed via jiwer with standard text normalization.
Key Takeaways
- • Deepgram Nova-3 leads on English WER (~5.2%) and streaming latency (~280ms final turn).
- • AssemblyAI Universal-1 has the best developer experience and bundled Audio Intelligence (summaries, sentiment, PII redaction, chapters).
- • OpenAI Whisper API remains best-in-class for multilingual (99 languages) but is batch-only and has no diarization.
- • Hyperscalers (AWS/GCP/Azure) are rarely cheapest or most accurate, but win when you need deep integration with their ecosystem.
- • Groq Whisper is the fastest batch option (LPU inference) and the cheapest hosted Whisper at ~$0.02/hr.
- • Self-hosted faster-whisper is the cheapest path at volume and the only option that gives you full data residency and offline capability.
- • No API reliably handles code-switching — Deepgram and AssemblyAI offer limited support (≤6 languages each).
- • Novascribe's July 2026 benchmark of 904 files across 16 datasets ranks AssemblyAI Universal-3.5 first on aggregate accuracy (7.0% avg WER), with OpenAI's GPT-4o Transcribe collapsing to 43.8% WER on financial calls despite winning short clean-audio benchmarks. Full data: Novascribe 2026 Benchmark section below.
Contents
Quick Picks by Use Case
| Use Case | API | Price | Why |
|---|---|---|---|
| Best overall, English production workloads | Deepgram Nova-3 | $0.0043–$0.0145/min | Lowest English WER, streaming + batch, strong diarization |
| Best developer experience | AssemblyAI | $0.12–$0.37/hr | Clean SDKs, Audio Intelligence add-ons, great docs |
| Best multilingual (99 languages) | OpenAI Whisper API | $0.006/min ($0.36/hr) | Largest language coverage, batch only |
| Best for accented English & EU residency | Speechmatics | From $0.30/hr | Enhanced model shines on accents; EU/UK hosting |
| Cheapest hosted Whisper | Groq Whisper | ~$0.02/hr | LPU inference, near real-time throughput, batch only |
| EU data residency, Whisper-compatible | Gladia | From €0.612/hr | FR-hosted, 100+ languages, diarization included |
| AWS-native pipeline | Amazon Transcribe | From $0.024/min | Call Analytics variant, custom vocab, S3-native |
| Microsoft stack / compliance | Azure AI Speech | ~$1/hr standard | 140+ languages, SOC2/HIPAA/FedRAMP options |
| Google Cloud shops | Google Speech-to-Text | $0.016–$0.024/min | Chirp v2 model, solid multilingual, V2 streaming |
| Need human fallback via API | Rev AI | $0.02/min AI | Same account covers AI async + human transcription |
| Budget Whisper-quality API | ElevenLabs Scribe | ~$0.22/hr | Newest entrant, 99 languages, aggressive pricing |
| Full data control / air-gapped | Self-hosted faster-whisper | Free + GPU compute | MIT license, ~$0.05–$0.15/hr cloud GPU |
APIs covered: Deepgram, AssemblyAI, OpenAI Whisper API, Speechmatics, Google Speech-to-Text, Azure AI Speech, Amazon Transcribe, Gladia, Rev AI, Groq Whisper, ElevenLabs Scribe, self-hosted faster-whisper.
What Changed in 2026
- • Deepgram Nova-3 launched with a redesigned acoustic model targeting call-center and noisy audio — the gap to Whisper on clean English is now within margin of error, and Deepgram wins clearly on phone/noisy audio.
- • AssemblyAI Universal-Streaming (2025) closed their real-time latency gap to Deepgram and added live Audio Intelligence.
- • OpenAI Realtime API is now the recommended path for conversational AI with streaming STT, but it is a separate billing and product from the Whisper API.
- • Groq began hosting Whisper large-v3 on LPU hardware at ~$0.02/hr — by far the cheapest hosted Whisper endpoint.
- • Gladia and Speechmatics emerged as the go-to EU-hosted options for GDPR-sensitive teams.
- • ElevenLabs Scribe entered the transcription API market with aggressive pricing.
- • Self-hosted Whisper matured: faster-whisper and whisper.cpp deliver 4–10× speedups, and Whisper accuracy is now a solved problem for most use cases.
Transcription API Pricing Reference (July 2026)
All prices are official list pricing for standard batch/streaming endpoints. Enterprise commitments, volume discounts, and reserved capacity can bring costs down 30–70%. Always confirm on the provider's pricing page before committing.
| API | Per-minute (List) | Per-hour | Free Tier | Model |
|---|---|---|---|---|
| Deepgram Nova-3 | $0.0043 (batch) / $0.0077 (stream) | $0.26 / $0.46 | $200 credit | Nova-3 |
| AssemblyAI Universal-1 | $0.0020 (batch) / $0.0025 (stream) | $0.12 / $0.15 | $50 credit + 185 free hrs | Universal-1 |
| OpenAI Whisper API | $0.006 | $0.36 | No | whisper-1 |
| Speechmatics Enhanced | ~$0.005 | $0.30 | 8 hrs/mo free | Enhanced |
| Groq Whisper large-v3 | ~$0.00033 | ~$0.02 | Rate-limited free tier | whisper-large-v3 |
| Google Speech-to-Text v2 | $0.016–$0.024 | $0.96–$1.44 | 60 min/mo | Chirp 2 |
| Azure AI Speech | $0.0167 | $1.00 | 5 hrs/mo | Standard |
| Amazon Transcribe | $0.024 (tier 1) | $1.44 | 60 min/mo × 12 mo | Standard |
| Gladia Whisper-Zero | ~€0.0102 | €0.61 | 10 hrs credit | Whisper-Zero |
| Rev AI | $0.02 (async) / $0.035 (stream) | $1.20 / $2.10 | 5 hrs/mo | Rev AI v3 |
| ElevenLabs Scribe | ~$0.0037 | ~$0.22 | Limited credits | Scribe v1 |
| Self-hosted Whisper (L4 GPU) | ~$0.001–$0.0025 | ~$0.05–$0.15 | Infra cost only | large-v3 / turbo |
Per-hour numbers are derived from list per-minute pricing (×60). Streaming endpoints are typically 20–80% more expensive than batch. Deepgram and AssemblyAI credits apply to both.
Novascribe's July 2026 Benchmark: 9 Models, 16 Datasets, 904 Files
We tested 9 major speech-to-text APIs against 904 audio files across 16 standard benchmarks in July 2026 — from Deepgram Nova-3 and AssemblyAI Universal-3.5 to OpenAI's Whisper-1, GPT-4o Transcribe, and GPT-4o Mini. This is the only cross-provider single-source WER comparison in the SERP for this query. Answer capsule and per-model results below.
Aggregate accuracy ranking (14 models, 16 datasets)
| Rank | Model | Overall avg WER | Multilingual avg | Note |
|---|---|---|---|---|
| 1 | AssemblyAI Universal-3.5 | 7.0% | 4.9% | Best of 9 tested |
| 2 | AssemblyAI Universal-2 | 8.1% | 6.4% | Solid legacy |
| 3 | Deepgram Nova-3 Multilingual | 8.2% | 8.2% | Cheaper than Nova-3 EN, weaker multilingual than U-3.5 |
| 4 | OpenAI Whisper-1 | 8.3% | 6.7% | Best OpenAI model for long-form |
| 5 | Deepgram Nova-3 English | 8.9% | n/a (English-only) | Tied best on English meetings |
| 6 | Deepgram hosted Whisper Large | 9.6% | 7.8% | 20-168s per file — dramatically slow |
| 7 | Deepgram Nova-2 | 10.5% | 9.7% | Legacy — Nova-3 EN barely beats it on English aggregate |
| 8 | GPT-4o Mini Transcribe | 12.0% | 5.7% | Best short-clip cost; collapses on long-form |
| 9 | GPT-4o Transcribe | 12.1% | 5.2% | Same collapse pattern at 2× the price |
Overall average WER = mean across all 16 English + multilingual datasets tested. Multilingual average = mean across FLEURS + CommonVoice for German, French, Spanish, Italian, Portuguese. Full per-language detail: our French, German, Spanish, Italian, and Portuguese pages.
The most important finding for API selection: the GPT-4o long-form collapse
GPT-4o Transcribe wins LibriSpeech (3.1% WER) but collapses on Earnings21 (43.8%) and TED-LIUM (27.1%). The same model that's near-perfect on 10-second audiobook clips cannot reliably transcribe a 30-minute earnings call. GPT-4o Mini has the identical failure pattern at half the price. If your users upload any content longer than a few minutes — meetings, podcasts, lectures, calls — route to Whisper-1 (9.7% on Earnings21) or AssemblyAI Universal-3.5 (12.4%), not the GPT-4o line.
Speed: seconds per file (short LibriSpeech clips vs long AMI meetings)
| Model | Short (~10s clips) | Long (~30min) |
|---|---|---|
| GPT-4o Mini Transcribe | 1.1s | 26.7s |
| GPT-4o Transcribe | 1.2s | 30.0s |
| OpenAI Whisper-1 | 1.6s | 45.5s |
| Deepgram Nova-2 | 2.3s | 29.5s |
| Deepgram Nova-3 English | 4.4s | 34.5s |
| AssemblyAI Universal-3.5 | 5.5s | 37.6s |
| Deepgram hosted Whisper Large | 20.4s | 168s |
Cost: total USD to run the full 904-file benchmark
| Model | Benchmark run cost | Per-minute rate |
|---|---|---|
| GPT-4o Mini Transcribe | $1.69 | $0.003/min |
| Deepgram Nova-3 Multilingual | $1.85 | Varies |
| AssemblyAI Universal-2 | $2.16 | Varies |
| AssemblyAI Universal-3.5 | $2.92 | Varies |
| GPT-4o Transcribe | $3.38 | $0.006/min |
| OpenAI Whisper-1 | $3.68 | $0.006/min |
| Deepgram hosted Whisper Large | $4.52 | Varies |
| Deepgram Nova-2 | $4.78 | Varies |
| Deepgram Nova-3 English | $7.40 | Varies |
Actual costs on our 904-file test set. Deepgram Nova-3 English is the most expensive despite sitting within ~0.4pp of Whisper-1 on English aggregate accuracy; GPT-4o Mini is the cheapest overall but only viable for short-form audio due to the collapse pattern above.
Methodology (this benchmark)
Test date: July 2026. 904 audio files across 16 standard benchmarks: LibriSpeech test-clean, AMI IHM, VoxConverse, Earnings21, TED-LIUM 3, GigaSpeech shard0, FLEURS (DE / FR / ES / IT / PT), CommonVoice 9 (DE / FR / ES / IT / PT), MLS-PT, plus 18 files of real Vexascribe production audio. All 9 models tested through official APIs with identical inputs. WER computed via jiwer with lowercase, punctuation-stripped normalization — the standard academic method. 95% bootstrap confidence intervals computed on datasets with ≥2 samples. No cherry-picking: all datasets included regardless of result; failures counted as errors.
Dataset licenses: LibriSpeech (CC BY 4.0), AMI (CC BY 4.0), VoxConverse (CC BY 4.0), Earnings21 (CC BY 4.0), TED-LIUM 3 (CC BY-NC-ND 3.0 — no transcripts reproduced), GigaSpeech (Apache 2.0), FLEURS (CC BY 4.0), CommonVoice (CC0), MLS (CC BY 4.0). Model versions tested reflect performance at time of test; providers update models frequently. For per-provider deep dives: our AssemblyAI, Deepgram, and Whisper accuracy pages.
English Accuracy Benchmarks (Word Error Rate)
Lower is better. Numbers below are composite estimates combining public vendor benchmarks (LibriSpeech test-clean, TED-LIUM, Switchboard) with our spot-checks on noisy and accented audio. Treat gaps below ~1 WER point as noise — they will flip based on your specific audio domain. For single-source cross-provider WER on identical audio, see our Novascribe 2026 Benchmark section above; for broader methodology see How accurate is Whisper?
| API | Clean English | Accented | Noisy | Phone (8kHz) |
|---|---|---|---|---|
| Deepgram Nova-3 | ~5.2% | ~7.1% | ~8.8% | ~9.4% |
| AssemblyAI Universal-1 | ~5.4% | ~7.6% | ~9.3% | ~10.1% |
| OpenAI Whisper large-v3 | ~5.5% | ~8.0% | ~10.5% | ~12.8% |
| Speechmatics Enhanced | ~5.8% | ~6.9% | ~9.0% | ~10.3% |
| Google Chirp v2 | ~6.1% | ~8.5% | ~11.0% | ~11.6% |
| Azure AI Speech | ~6.5% | ~9.0% | ~11.5% | ~12.0% |
| Amazon Transcribe | ~7.0% | ~9.5% | ~11.8% | ~11.2% (Call Analytics) |
| Gladia Whisper-Zero | ~5.6% | ~8.2% | ~10.8% | ~13.0% |
| Rev AI v3 | ~6.3% | ~8.9% | ~10.7% | ~11.0% |
| ElevenLabs Scribe | ~5.7% | ~8.4% | ~10.9% | ~12.4% |
Reality check: For clean English podcast or meeting audio, all top APIs are within 1–2 WER points. Pick based on latency, diarization, and pricing. The gap opens on phone/noisy audio, where Deepgram Nova-3, Speechmatics Enhanced, and AssemblyAI clearly outperform generic Whisper.
Streaming Latency
For interactive products (voice assistants, live captions, conversational AI), latency matters more than raw WER. “First token” is how fast you get any text back; “final turn” is how fast you get the finalized transcript after the speaker stops.
| API | First Token | Final Turn | Notes |
|---|---|---|---|
| Deepgram (Nova-3 Streaming) | ~150ms | ~280ms | Purpose-built real-time engine |
| AssemblyAI Universal-Streaming | ~200ms | ~400ms | Released 2025, sub-500ms target |
| Speechmatics RT | ~180ms | ~450ms | Strong on accented speech |
| Azure Speech SDK | ~250ms | ~600ms | WebSocket or SDK streaming |
| Google Speech v2 Streaming | ~300ms | ~700ms | gRPC streaming, Chirp v2 batch-only |
| Rev AI Streaming | ~350ms | ~800ms | Adequate for meetings, not conversational AI |
| OpenAI Whisper API | N/A (batch only) | ~4–15s for 1-min audio | No streaming endpoint; use Realtime API for conversational |
| Groq Whisper | N/A (batch only) | ~1–3s for 1-min audio | Fastest batch throughput (LPU) |
| Self-hosted faster-whisper | N/A (batch, but can chunk) | Depends on GPU / chunking strategy | Roll your own streaming with 30s windows |
Numbers measured from US-East clients on stable networks. Your real-world latency will depend on region, audio codec, and SDK buffering defaults. For sub-second round-trip conversational AI, Deepgram + OpenAI Realtime is the common pairing.
Feature Matrix
| API | Streaming | Diarization | Languages | Code-switch | Translation | Customization | EU Residency |
|---|---|---|---|---|---|---|---|
| Deepgram | ✓ | ✓ | 40 | ✓ | ✗ | Keyterm boosting | Optional EU region |
| AssemblyAI | ✓ | ✓ | 99 | ✓ | ✗ | Word boost, Audio Intelligence | US default, EU via enterprise |
| OpenAI Whisper API | ✗ | ✗ | 99 | partial | ✓ | Prompt parameter | Enterprise EU residency |
| Speechmatics | ✓ | ✓ | 55 | ✗ | ✓ | Custom dictionary | EU/UK native |
| Google Speech v2 | ✓ | ✓ | 125 | ✗ | ✗ | Model adaptation | EU regions available |
| Azure Speech | ✓ | ✓ | 140 | ✗ | ✓ | Custom Speech model | EU regions, sovereign cloud |
| Amazon Transcribe | ✓ | ✓ | 100 | partial | ✗ | Custom vocab, custom LM | EU regions available |
| Gladia | ✓ | ✓ | 100 | ✗ | ✓ | Prompt/vocabulary | FR-hosted native |
| Rev AI | ✓ | ✓ | 37 | ✗ | ✗ | Custom vocab | US default |
| Groq Whisper | ✗ | ✗ | 99 | partial | ✓ | Prompt parameter | US only |
| ElevenLabs Scribe | ✗ | ✓ | 99 | ✗ | ✗ | Speaker labels | US default |
| faster-whisper (OSS) | ✗ | via pyannote | 99 | partial | ✓ | Initial prompt, LoRA | Self-hosted — you decide |
Detailed Reviews
Each review below covers accuracy, latency, pricing model, SDK quality, and the audio workloads where each API is the right or wrong choice.
1. Deepgram Nova-3
Best OverallLowest-latency production API with strongest English WER
Deepgram built its own end-to-end ASR stack from scratch (not Whisper). Nova-3 is purpose-built on call-center and conversational audio, which is why it beats Whisper on noisy and phone-quality audio by 2–4 WER points. Streaming latency is the lowest in the industry (~280ms final turn), and diarization is solid out of the box. SDKs cover Node, Python, .NET, Go, and Rust, with a well-documented WebSocket streaming protocol. The main trade-off is language coverage — 40 languages vs 99 for Whisper.
Novascribe July 2026 benchmark: Nova-3 English measured 12.3% average English WER across our tested set — within ~0.4pp of OpenAI Whisper-1 (11.9%) and AssemblyAI Universal-2 (11.9%) on aggregate. Strongest on English meetings (AMI 20.9%, tied best of 9 models tested). But dramatically underperforms on non-English: FLEURS German 8.0% vs AssemblyAI Universal-3.5's 2.6%. Nova-3 Multilingual improved cost but not accuracy: 8.2% aggregate multilingual vs Universal-3.5's 4.9%. See our Deepgram accuracy page for the full per-language breakdown.
Best For
- •Real-time English products
- •Call-center/phone audio
- •High-volume streaming at scale
Pros
- ✓Lowest streaming latency in tested set
- ✓Best English WER in noisy/phone audio
- ✓Competitive batch pricing ($0.0043/min)
- ✓Strong diarization and keyterm boosting
Cons
- ✗Only 40 languages (vs 99 Whisper)
- ✗No built-in translation
- ✗EU region is a request-only option
2. AssemblyAI
Best Developer DXClean SDKs + bundled Audio Intelligence (summaries, sentiment, PII)
AssemblyAI's Universal-1 model reaches parity with Whisper on clean English and their 2025 Universal-Streaming release closed the real-time latency gap to Deepgram. The differentiator is Audio Intelligence: auto chapters, summarization, sentiment, entity detection, PII redaction, and topic detection all available as flags on the same request. If you need transcription plus LLM-style post-processing without running your own pipeline, nothing else is this integrated. SDKs are idiomatic in all major languages and the docs are consistently rated the best in the category.
Novascribe July 2026 benchmark: AssemblyAI Universal-3.5 achieved the best aggregate accuracy of 9 tested models (7.0% average WER across 16 datasets). Best-in-class on multilingual: 4.9% average across German, French, Spanish, Italian, Portuguese. Remarkable German result: 0.9% WER on accented CommonVoice-DE, actually better than clean FLEURS-DE (2.6%). Italian at 1.4% on FLEURS is the strongest single-language result of any model tested. Weakness: accented Portuguese (CommonVoice-PT: 12.1%) and French (CommonVoice-FR: 9.9%) — test with your own audio if those matter. Full data: our AssemblyAI accuracy page.
Best For
- •Meeting assistants / note-takers
- •Content workflows needing summaries
- •Teams that value SDK polish
Pros
- ✓Best-in-class SDKs and docs
- ✓Bundled summaries, sentiment, PII, chapters
- ✓Batch pricing ($0.12/hr) extremely competitive
- ✓99 languages for Universal-1
Cons
- ✗Audio Intelligence features stack extra cost
- ✗EU residency requires enterprise contract
- ✗Streaming latency slightly behind Deepgram
3. OpenAI Whisper API
Best Multilingual99 languages, dead-simple API, batch only, no diarization
OpenAI's hosted Whisper API is the fastest way to get 99-language transcription into a product. The API takes audio + optional prompt and returns text, SRT, or VTT — no tuning, no SDK beyond the standard OpenAI client. The catches are real: there is no streaming endpoint (use the Realtime API for conversational audio), no built-in diarization (pair with WhisperX or pyannote), and no word-level confidence in the standard response. For batch multilingual transcription of uploaded files, it's hard to beat. For interactive products, pick Deepgram or AssemblyAI.
Novascribe July 2026 benchmark — critical distinction: Whisper-1 (the older whisper-1 endpoint) held 9.7% WER on financial earnings calls (Earnings21) while GPT-4o Transcribe collapsed to 43.8% and GPT-4o Mini to 44.2% on the same audio. Same pattern on TED-LIUM long-form speech: Whisper-1 5.0% vs GPT-4o 27.1%. For long-form or noisy content, use Whisper-1. For short clean voice memos and short conversational clips, GPT-4o Mini offers similar-or-better accuracy at half the price ($0.003/min vs $0.006/min). Full breakdown: our Whisper accuracy page.
Best For
- •Multilingual file transcription
- •Teams already on OpenAI
- •Prototype/MVP fast-path
Pros
- ✓99 languages out of the box
- ✓Trivial integration via existing OpenAI SDK
- ✓Built-in translation (any lang → English)
Cons
- ✗No streaming (batch only)
- ✗No diarization or speaker labels
- ✗25 MB file limit per request
- ✗No EU residency on default API
4. Speechmatics
Best for Accents + EUAccent-robust ASR with native UK/EU hosting
Speechmatics is a UK company whose Enhanced model has consistently outperformed competitors on accented English (Indian, African, Caribbean) in independent benchmarks. Native EU/UK hosting with signed DPAs makes it a common pick for GDPR-sensitive teams who can't wait on enterprise paperwork from US providers. Streaming and batch are both first-class, 55 languages supported, and translation is built in. Pricing is middle-of-pack but transparent.
Best For
- •Accented English (broadcast, global calls)
- •UK/EU compliance teams
- •Broadcast media workflows
Pros
- ✓Strongest accented-English accuracy
- ✓Native EU/UK data residency
- ✓Streaming + batch in one API
- ✓Built-in translation
Cons
- ✗Pricier than Deepgram/AssemblyAI at scale
- ✗Smaller SDK ecosystem
- ✗55 languages vs 99 for Whisper
5. Gladia
Best EU Whisper APIFR-hosted Whisper-compatible API with diarization included
Gladia is a French provider offering a hardened Whisper pipeline (“Whisper-Zero”) with word-level timestamps, diarization, and translation included as flags. Hosting is FR-native with signed DPAs — the most painless path to a GDPR-compliant Whisper API. Pricing is higher than raw Whisper but includes diarization and post-processing you'd otherwise bolt on yourself.
Best For
- •EU SaaS products
- •Teams that want Whisper + diarization
- •French-market media/meeting apps
Pros
- ✓FR/EU-hosted by default
- ✓Diarization + translation bundled
- ✓100+ languages via Whisper
Cons
- ✗More expensive than raw Whisper
- ✗Streaming is newer, less mature than Deepgram
6. AWS Transcribe / Google Speech / Azure Speech
Best for Cloud-nativeHyperscaler APIs — ecosystem depth trumps raw accuracy
The three hyperscalers are rarely the cheapest or most accurate option, but they win when you need deep integration with the rest of the cloud — S3 lifecycle rules, Google Cloud Storage triggers, Azure Logic Apps, compliance certifications already negotiated. Amazon Transcribe Call Analytics is specifically strong for AWS Connect contact centers. Google's Chirp v2 is competitive on multilingual. Azure Speech covers 140+ languages and supports sovereign-cloud deployments. If your architecture lives inside one of these clouds, the integration savings often outweigh a small accuracy gap.
Best For
- •Teams already in AWS/GCP/Azure
- •Contact centers (AWS Connect)
- •Compliance-heavy enterprises
Pros
- ✓Native integration with cloud storage/events
- ✓Existing compliance envelopes (SOC2, HIPAA, FedRAMP)
- ✓Regional deployment and sovereign-cloud options
Cons
- ✗3–10× more expensive than Deepgram/AssemblyAI
- ✗Lower accuracy on noisy/phone audio
- ✗Heavier SDKs and IAM overhead
7. Self-hosted Whisper (faster-whisper)
Best for Data ControlFree, 99 languages, full data residency on your infrastructure
faster-whisper (CTranslate2 backend) and whisper.cpp (GGML) are the two production-grade Whisper reimplementations. Expect 4–10× speedup over the reference OpenAI implementation on the same hardware. A single L4 or A10G handles ~100× real-time with large-v3, making self-hosting the cheapest option at >500 hrs/month. You get full data control, offline capability, and the ability to fine-tune on domain audio. You also own the ops: GPU autoscaling, queue management, retries, and monitoring. Pair with pyannote.audio for diarization and you have a feature-complete pipeline.
Best For
- •Volume > 500 hrs/month
- •Strict data residency / air-gapped
- •Domain fine-tuning needs
Pros
- ✓Cheapest at volume (pennies per hour)
- ✓Full data residency, offline capable
- ✓99 languages, MIT license
- ✓Fine-tune on your domain audio
Cons
- ✗You own GPU ops, autoscaling, retries
- ✗Streaming needs custom chunking
- ✗Diarization is a separate pipeline
Failure Modes to Know (From Our July 2026 Benchmark)
Every provider's marketing page tells you what its model does well. These are the failure modes we measured in our own testing that we would want to know before shipping. Each is one benchmark result, not the last word — treat them as prompts to test with your own audio.
GPT-4o long-form collapse
GPT-4o Transcribe: 3.1% WER on LibriSpeech short clips, 43.8% WER on Earnings21, 27.1% on TED-LIUM. GPT-4o Mini has the identical failure pattern at half the price. Do not use either for any audio longer than ~2 minutes: podcasts, meetings, earnings calls, lectures, interviews all collapse. Whisper-1 holds at 9.7% on the same Earnings21 audio.
Nova-3 Multilingual disappointment
Deepgram Nova-3 Multilingual's 8.2% aggregate multilingual WER trails AssemblyAI Universal-3.5 (4.9%) and even OpenAI Whisper-1 (6.7%). On Italian FLEURS, Nova-3 English (2.8%) actually beat Nova-3 Multilingual (3.6%) — the reverse of its positioning. For serious multilingual production audio, choose AssemblyAI U-3.5 or Whisper-1 instead of Nova-3 ML.
Deepgram hosted Whisper Large severe latency
Deepgram's hosted OpenAI Whisper endpoint returned 168 seconds per AMI meeting file vs 26–45 seconds for every direct alternative — a 4–8× latency cost. Its one accuracy win in our benchmark (CommonVoice French, 6.2% WER) is genuinely notable, but the throughput penalty rules it out for most production workflows. Prefer direct OpenAI Whisper-1 or AssemblyAI Universal-3.5.
How to Pick
Ignore marketing. Start from your constraints:
1. Streaming or batch?
If you need sub-second transcripts as the user speaks, you are choosing between Deepgram, AssemblyAI, Speechmatics, and Azure. Whisper API is off the table for streaming.
2. English-only or multilingual?
English-only → Deepgram Nova-3 wins on accuracy + price. Multilingual at 10+ languages → Whisper-based (OpenAI, Gladia, Groq, or self-hosted) for 99-language coverage.
3. Data residency requirements?
EU required → Speechmatics or Gladia out of the box. Strict (no third-party at all) → self-hosted Whisper on your infrastructure.
4. What volume?
<50 hrs/month → hosted API, pick on DX. 50–500 hrs/month → Deepgram or AssemblyAI with committed pricing. >500 hrs/month → self-hosted faster-whisper starts winning on TCO.
5. Do you need diarization, summaries, or sentiment?
Yes → AssemblyAI ships it in one request. Otherwise plan for a separate pipeline (Whisper + pyannote + an LLM).
Always benchmark on your own audio before committing. Free tiers from Deepgram, AssemblyAI, and Gladia cover enough minutes to run a real evaluation. Don't trust any provider's headline WER — it was measured on audio that isn't yours.
When You Don't Need an API
Developers sometimes reach for a transcription API when a hosted product would solve their actual problem faster and cheaper:
- • Users upload files and want transcripts: a hosted UI like VexaScribe, TurboScribe, or Happy Scribe handles upload, processing, editing, and export without you building any of it.
- • Internal team needs meeting notes: Otter, Fireflies, or a meeting bot is faster than integrating any API.
- • One-off bulk transcription project: TurboScribe unlimited or VexaScribe at $0.20–$0.60/hr is cheaper and faster than wiring up an API.
If you fall into any of those buckets, skip the API and use a hosted tool. If you're embedding transcription into a product, proceed with the API comparison above. For context on choosing between hosted products, see best transcription software 2026.
Note on VexaScribe: We are a hosted transcription product, not a transcription API provider. We recommend the APIs above purely on their merits for developers building speech-to-text into their own products. If you just need transcripts from audio you or your users upload, VexaScribe's UI uses the Whisper large-v3 model too — without the integration work.
Deepgram vs AssemblyAI vs Whisper — Head-to-Head
The three APIs everyone actually shortlists. Same audio, three different tradeoffs.
| Dimension | Deepgram Nova-3 | AssemblyAI Universal-1 | OpenAI Whisper API |
|---|---|---|---|
| Batch price | $0.0043/min ($0.26/hr) | $0.12/hr | $0.006/min ($0.36/hr) |
| Streaming | Yes, ~$0.0077/min | Yes, ~$0.15/hr | No (Realtime API is separate) |
| English WER (clean) | ~5.2% | ~5.4% | ~6.0% |
| Noisy/phone audio | Clear leader | Strong | Weakest of the three |
| Languages | 36+ | 99 | 99 |
| Diarization | Included | Included | Not included |
| Audio Intelligence add-ons | Summaries, sentiment (paid) | Broadest set (summaries, sentiment, PII redaction, chapters) | None — you build it |
Deepgram vs Whisper
Deepgram wins on production English (lower WER on noisy/phone audio, sub-300ms streaming latency, and is the cheaper hosted option at $0.0043/min batch vs Whisper API's $0.006/min). Whisper wins on language coverage (99 vs Deepgram's 36+) and multilingual accuracy. Pick Deepgram for English voice products; pick Whisper API when you need broad language support and can accept batch-only.
Deepgram vs AssemblyAI
Deepgram wins on price and raw streaming latency. AssemblyAI wins on developer experience and pre-built Audio Intelligence (summaries, sentiment, PII redaction, chapters, entity detection) — if you'd have to build those on top of Deepgram anyway, AssemblyAI ends up cheaper. Pick Deepgram for high-volume raw transcription; pick AssemblyAI when the downstream analysis matters as much as the transcript.
AssemblyAI vs Whisper
AssemblyAI is a full transcription platform; Whisper API is a raw ASR endpoint. AssemblyAI wins if you want summaries, sentiment, redaction, and diarization from one call. Whisper wins on price for pure transcription ($0.006/min vs $0.12/hr) and on multilingual breadth. Whisper has no diarization — you'd need WhisperX or pyannote on top.
Deepgram Alternatives & Competitors
If you're specifically shopping Deepgram alternatives, the shortlist depends on what you're replacing:
- • Cheaper hosted Whisper → Groq at $0.02–$0.04/hr (5–9× cheaper than Deepgram batch and 9× cheaper than OpenAI's Whisper API for the same underlying model)
- • Better multilingual coverage → OpenAI Whisper API, Gladia, or ElevenLabs Scribe (all 99 languages vs Deepgram's 36)
- • EU data residency → Speechmatics (UK), Gladia (FR), or self-hosted Whisper
- • Bundled Audio Intelligence → AssemblyAI (summaries, sentiment, PII redaction, chapters in one call)
- • Full data control / offline → self-hosted faster-whisper on your own GPU (~$0.05–$0.15/hr compute-only)
- • Hyperscaler stack → AWS Transcribe, Google Cloud Speech-to-Text, or Azure AI Speech if you're already committed to that cloud
Is Deepgram free?
Deepgram offers $200 in free credits on signup — roughly 775 hours of Nova-3 batch transcription at $0.0043/min, or ~430 hours of Nova-3 streaming at $0.0077/min. No credit card required to activate. That's enough free capacity to build and validate most integrations before paying anything. There is no perpetual free tier — once credits run out, pay-as-you-go begins.
Is AssemblyAI free?
AssemblyAI offers $50 in free credits on signup (~415 hours of Universal-1 batch at $0.12/hr). No card required. Same pattern as Deepgram: enough to build and test, no perpetual free tier.
Cheapest transcription API?
Cheapest hosted: Groq Whisper at $0.02–$0.04/hr (Whisper large-v3 on Groq's LPU hardware). Cheapest at volume: self-hosted faster-whisper on your own GPU, ~$0.05–$0.15/hr compute-only. Between those: Deepgram Nova-3 batch ($0.26/hr) and AssemblyAI Universal-1 batch ($0.12/hr). OpenAI's Whisper API sits at $0.36/hr — the most expensive way to run Whisper by a wide margin.
Free transcription API options?
Truly free (unlimited): self-hosted Whisper — MIT license, MP4/MP3/audio direct, but you pay for compute. Free with signup credits: Deepgram ($200), AssemblyAI ($50), OpenAI ($5 initial credit on new accounts). Groq offers a free tier (2,000 speech-to-text requests/day) — the most generous perpetual free tier of any hosted transcription API.
Frequently Asked Questions
Frequently Asked Questions
What's the cheapest transcription API in 2026?
Self-hosted OpenAI Whisper is free (you pay only for compute). Among hosted APIs, Deepgram Nova-3 ($0.0043/min ≈ $0.26/hr) and Groq's hosted Whisper ($0.02/hr) are the cheapest. OpenAI's Whisper API sits at $0.006/min ($0.36/hr). AssemblyAI Universal-1 is $0.12/hr batch. Rev AI and Google Speech-to-Text sit higher at $0.30–$1.44/hr depending on features.
Which transcription API has the lowest latency for real-time?
Deepgram (sub-300ms streaming), Speechmatics (sub-500ms), and AssemblyAI Universal-Streaming (sub-400ms) lead for real-time. OpenAI Whisper API is batch-only — no true streaming endpoint. For sub-second latency you need a purpose-built streaming engine, not Whisper.
Is OpenAI's Whisper API the most accurate?
Not anymore. Whisper large-v3 leads in multilingual coverage (99 languages), but on clean English audio Deepgram Nova-3 and AssemblyAI Universal-1 match or beat it (WER ≈5%). On noisy or accented audio, Deepgram and Speechmatics typically outperform Whisper. For non-English, Whisper remains best-in-class.
Does OpenAI have a streaming Whisper API?
No. OpenAI's Whisper API is batch-only. The Realtime API (GPT-4o with audio) supports streaming speech-to-text but is billed differently (~$0.06/min input audio) and optimized for conversational AI, not pure transcription. For streaming ASR at scale, use Deepgram, AssemblyAI, or Speechmatics.
Which API has the best speaker diarization?
AssemblyAI and Deepgram both offer strong diarization (2–10 speakers, ~90% accuracy). Pyannote (open source) is the academic benchmark. OpenAI Whisper API does NOT include diarization — you must run WhisperX or pyannote separately. Speechmatics also ships solid diarization with its Enhanced model.
Can I self-host Whisper for production workloads?
Yes. Whisper is MIT-licensed and runs on a single GPU. For production, use faster-whisper (CTranslate2) or whisper.cpp — 4–10× faster than the reference implementation. A single A10G or L4 GPU handles ~100× real-time with large-v3. Expect ~$0.05–$0.15/hr in cloud GPU cost — cheaper than most hosted APIs at volume.
Does any API support code-switching (mixed languages)?
AssemblyAI and Deepgram both support code-switching on a limited subset of languages (≤6 each). Most APIs lock you to one language per request. Whisper technically detects language shifts but outputs degrade on true code-switching. No API solves this perfectly — benchmark on your actual audio.
Are there GDPR-compliant transcription APIs with EU data residency?
Yes. Speechmatics (UK), Amberscript (NL), and Gladia (FR) offer EU data residency and signed DPAs. AWS Transcribe and Azure Speech let you pick an EU region. OpenAI offers EU data residency for enterprise contracts but not on the default Whisper API. For strict GDPR, self-hosted Whisper eliminates the question entirely.
Which API is best for noisy call-center audio?
Deepgram Nova-3 (purpose-built on contact-center data), AssemblyAI Universal-1, and Speechmatics Enhanced consistently outperform Whisper on 8kHz telephony audio with noise and overlap. For call centers specifically, Deepgram's Nova-3 phonecall model is the standard pick.
Do I need a transcription API if my users just want to upload files?
Probably not. If your product is a consumer-facing transcription tool, a hosted UI like VexaScribe handles upload, processing, editing, and export without you touching an API. APIs make sense when you're embedding transcription into a larger product (meeting assistants, compliance tooling, media pipelines) — not when a finished UI would do.
Which transcription API is most accurate in 2026?
AssemblyAI Universal-3.5, by our July 2026 benchmark of 904 audio files across 16 datasets. It ranked first among 9 tested speech-to-text models with 7.0% average WER overall, and best on multilingual (4.9% average across German, French, Spanish, Italian, Portuguese). Deepgram Nova-3 English tied for best on English meeting audio (AMI 20.9% WER) but underperformed on non-English. OpenAI Whisper-1 was the strongest OpenAI transcription model at 8.3% overall — crucially, it held 9.7% WER on financial earnings calls where GPT-4o Transcribe collapsed to 43.8%. For maximum aggregate accuracy: AssemblyAI. For English speed: Deepgram. For 99-language batch: Whisper-1.
Is GPT-4o Transcribe good for production transcription APIs?
Only for short (<2 minute), clean audio. Novascribe's July 2026 benchmark surfaced a critical failure pattern: GPT-4o Transcribe wins clean short-clip benchmarks (LibriSpeech 3.1% WER) but collapses on long or noisy audio: 43.8% WER on Earnings21 financial calls, 27.1% on TED-LIUM long prepared speech, 40.9% on AMI meetings. GPT-4o Mini has the identical failure pattern at half the cost. If your users upload voice memos, customer service snippets, or short conversational clips, GPT-4o Mini is the cheapest good option ($0.003/min). For anything longer — podcasts, meetings, earnings calls, lectures, interviews — use OpenAI Whisper-1 (9.7% WER on Earnings21) or AssemblyAI Universal-3.5 (12.4%) instead.
Should I use Deepgram Nova-3 or Nova-3 Multilingual for non-English?
Neither if accuracy matters most. Novascribe's July 2026 benchmark measured Deepgram Nova-3 Multilingual at 8.2% aggregate multilingual WER — worse than OpenAI Whisper-1 (6.7%) and significantly worse than AssemblyAI Universal-3.5 (4.9%). Nova-3 Multilingual costs less but doesn't fulfill its multilingual positioning: on Italian FLEURS, Nova-3 English (2.8%) actually beat Nova-3 Multilingual (3.6%). For serious multilingual production audio, use AssemblyAI Universal-3.5 or OpenAI Whisper-1. One notable anomaly: Nova-3 English surprisingly won accented CommonVoice Portuguese at 6.2% WER — worth testing if your Portuguese audio is heavily accented Brazilian.