Transcribe a recording accurately

Lowest measured error rate, diarisation included, priced like a commodity.

Use it when

You have a finished recording — interview, podcast, deposition, lecture — and you want the best transcript available without running infrastructure.

Cost

$0.22/hr batch ($3.67 per 1,000 min). Realtime $0.39/hr.Scribe v2 $0.22/hr; Scribe v2 Realtime $0.39/hr. Add-ons billed separately: entity detection $0.070/hr, keyterm prompting $0.050/hr. No standalone free tier on the API page — the Starter plan at $6/month includes 4 hours 30 minutes of Scribe v2. Files to 3 GB and 10 hours per job.

Good at

Artificial Analysis ranks it first on AA-WER v2 at 2.2% WER, ahead of Microsoft's MAI-Transcribe-1.5 (2.4%) and Google's Gemini 3.5 Transcribe (2.6%), and it is cheaper than both at $3.67 per 1,000 minutes against $6.00 and $5.00. Diarisation up to 32 speakers and word-level timestamps are in the base price, not bolted on. 90+ languages.

The catch

ElevenLabs will not tell you what the model is. The Scribe v2 launch post claims 'the lowest word error rate recorded on industry-standard benchmarks' and prints a FLEURS chart with no figures on it and no named competitors — you are trusting a third-party leaderboard, not the vendor. There is no self-hosted or on-prem option of any kind, so if your audio is legally confined to your own infrastructure this is a dead end. Keyterm prompting, the feature you will want the moment your recording contains product names or drug names, costs extra per hour.

Wrong for

Anyone with data-residency or air-gap constraints — go to Whisper self-hosted at rank 4 or Speechmatics' container at rank 5. Also wrong if you need streaming for a voice agent, where AssemblyAI's realtime tier is the stronger buy.

Best on genuinely bad audio. Diarisation costs extra.

Use it when

The recording is rough — phone audio, a laptop mic across a boardroom table, clipping, heavy accents — and you have tried a cheap model already and thrown the output away.

Cost

$0.21/hr async, plus $0.02/hr for speaker diarisation.Universal-3.5 Pro async $0.21/hr; Universal-2 async $0.15/hr. Streaming: Universal-3.5 Pro Realtime $0.45/hr, Universal-Streaming $0.15/hr (multilingual also $0.15/hr). Sync API $0.45/hr. Speaker diarisation is a $0.02/hr add-on on both async and streaming. Free tier covers up to 185 hours of pre-recorded transcription per AssemblyAI's own pricing page. Volume pricing by sales contact.

Good at

Degraded audio. AssemblyAI reports Universal-3.5 Pro led all 31 models on Coval's independent benchmark at 3.4% overall WER, and topped the field specifically on clipping and far-field audio — the two conditions that wreck real meeting recordings. The free tier is the most generous here by a wide margin: 185 hours is enough to transcribe a whole back catalogue before you pay anything.

The catch

Diarisation is a separate line item, so the $0.21/hr you budgeted is actually $0.23/hr the moment you need speaker labels, and every comparison table you have seen quotes the number without it. The Coval result is reported by AssemblyAI on AssemblyAI's blog — Coval's own public write-up declines to publish per-provider rankings, so you cannot check the 3.4% independently from the source. And the sync API, which is what you will reach for when you want a transcript back in one HTTP call rather than polling, is $0.45/hr — more than double the async rate.

Wrong for

Clean single-speaker audio at volume, where Deepgram or Speechmatics Melia cost a third as much for a transcript you will not be able to tell apart. Also wrong if the audio cannot leave your infrastructure.

Fastest hosted transcription, diarisation free, accuracy a step behind.

Use it when

You are transcribing thousands of hours and throughput matters more than the last percentage point — call-centre archives, media libraries, anything where the transcript feeds search rather than a human reader.

Cost

$0.0043/min pre-recorded English (~$0.26/hr). Diarisation included.Pay-as-you-go pre-recorded: Nova-3 monolingual $0.0043/min, multilingual $0.0052/min. Growth plan: $0.0036/min and $0.0043/min. Streaming: Nova-3 monolingual $0.0048/min, multilingual $0.0058/min, Flux English $0.0065/min. Speaker diarisation is included on pre-recorded, but costs $0.0020/min extra on streaming. $200 promotional credit for new accounts. Self-hosted via Docker, Podman or Kubernetes requires a Deepgram Enterprise Plan; Amazon SageMaker deployment is billed through AWS without an enterprise contract.

Good at

Speed and unit economics. Artificial Analysis clocks Nova-3 at 530.6x realtime, the fastest non-streaming model on their board — a two-hour recording comes back in well under a minute. Diarisation is genuinely included on pre-recorded, unlike AssemblyAI. And it is the only major hosted vendor here with a documented self-hosted path that does not start with a sales call, via SageMaker.

The catch

It is not the accuracy leader and Deepgram's own pricing page is where the good news stops — the product page for Nova-3 returned a 404 when I went looking for its WER claims, so I have no vendor-stated accuracy figure to attribute at all. On the independent AA board it does not appear in the top five by WER. That gap shows up as exactly the errors that cost you time: proper nouns, alphanumeric identifiers, overlapping speech. The $200 credit is generous but expires, and Growth pricing requires a monthly commitment.

Wrong for

A single important recording where every word matters — pay the extra 4p an hour for Scribe v2 or Universal-3.5 Pro. Also wrong for Docker-based on-prem unless you are prepared for an enterprise contract.

Free, offline, 99 languages. Hallucinates, and you own the GPU.

Use it when

The audio legally cannot leave your infrastructure — medical, legal, HR, defence — or you have enough volume that per-minute API pricing has stopped being rounding error.

Cost

Free. Apache 2.0 weights, MIT runtime. You pay for GPU only.Model weights Apache 2.0, 1.55B parameters, 99 languages. faster-whisper is MIT-licensed and claims up to 4x faster than openai/whisper at the same accuracy with less memory (13 minutes of audio: 2m23s on openai/whisper versus 1m03s on faster-whisper, large-v2 FP16 on GPU). WhisperX adds word-level timestamps and speaker diarisation via pyannote, which has its own model-access gating. If you would rather rent than own, Groq hosts it at $0.111/hr for large-v3 and $0.04/hr for large-v3-turbo — but that is a hosted API again, with a 25 MB file cap on the free tier and 100 MB on dev.

Good at

Nothing else on this list runs with the network cable pulled out. 99 languages is the broadest coverage here — Parakeet does 25, Scribe does 90+. The ecosystem is the real asset: faster-whisper, WhisperX, whisper.cpp, distil-whisper and a decade of fine-tunes all take the same weights, so you can trade accuracy for speed without changing vendors. And the marginal cost of the ten-thousandth hour is the same as the first.

The catch

OpenAI's own model card documents it: the model can generate text 'not actually spoken in the audio input', and the sequence-to-sequence architecture 'makes it prone to generating repetitive texts'. On day three you will find a paragraph of plausible, fluent, entirely invented speech in the middle of a silence, and no confidence score flagged it. It also ships with no diarisation — you bolt on WhisperX and pyannote yourself, and now you are maintaining a three-model pipeline. Performance is explicitly uneven across accents, dialects and low-resource languages. Budget a week, not an afternoon.

Wrong for

Anyone in a hurry with a single recording, or a team with no GPU and no appetite for pipeline maintenance. Pay $0.22/hr and move on. Also wrong if you need low-latency streaming.

Cheapest credible hosted rate, plus a container you can actually deploy.

Use it when

You need on-prem or private-cloud deployment with a support contract behind it, or you are transcribing at volume in many languages and want the price floor.

Cost

Batch from $0.129/hr (Melia 1). Enhanced $0.40/hr. Container on Enterprise.Batch: Melia 1 $0.129/hr, Standard $0.24/hr, Enhanced $0.40/hr. Real-time: Standard $0.24/hr, Enhanced $0.43/hr. Add-ons: translation $0.65/hr, summaries $0.12/hr, chapters $0.40/hr. Diarisation is listed as a core speech-to-text feature across tiers, not an add-on. $100 credit to start, no card. Pro tier gives 50 concurrent real-time sessions and 10 file jobs/second; Free and Pro are SaaS-only across US, EU or Australia regions. Enterprise unlocks Private Cloud, Container, Virtual Appliance and On-Device — pricing is sales-gated, so check current pricing.

Good at

The deployment spread. It is the only vendor here offering a supported container and virtual appliance alongside a public hosted price list, which matters when compliance says on-prem but your team has no interest in babysitting a Whisper pipeline. $0.129/hr on Melia 1 is roughly half the going hosted rate. 55+ languages, diarisation included, and a $100 no-card credit that is a real trial rather than a demo.

The catch

The $0.129/hr headline is the cheapest model, not the good one — the tier you will end up on after listening to the output is Enhanced at $0.40/hr, three times the advertised floor and nearly double Scribe v2. Speechmatics publishes no WER figure I could verify on their pricing page, and they appear in AssemblyAI's competitor list without a number attached, so the accuracy case rests on reputation rather than a benchmark I can point you at. On-prem pricing is entirely behind a sales conversation — I could not verify it.

Wrong for

A solo builder with one recording and no procurement process. The trial is fine, but the value here is the enterprise deployment path, and you are not buying that.