Best AI Transcription Tools 2026: Price Per Hour, Diarisation and Languages Compared

What changed since Whisper

The single model has become a lineup. Alongside the original Whisper endpoint, OpenAI's pricing page now lists gpt-transcribe, gpt-4o-transcribe and gpt-4o-mini-transcribe for batch work. It adds gpt-4o-transcribe-diarize for speaker attribution, plus gpt-live-transcribe and gpt-realtime-whisper for streaming at $0.017 per minute. The newer models are billed on tokens with a per-minute equivalent shown for guidance, which matters when you are forecasting a large workload.

Self-hosting, and when it is worth it

The open Whisper weights are still the reason this entry ranks first for anyone with a compliance constraint. Running large-v3 locally needs roughly 10GB of VRAM and processes audio several times faster than real time on a modern GPU, and smaller variants trade accuracy for lower hardware requirements. For legal depositions, clinical consultations and recorded financial calls, self-hosting means no audio leaves your tenant, and no data processing agreement has to cover it. Pair it with a diarisation library, because the open model does not attribute speakers on its own.

The ecosystem effect

Whisper's release raised the accuracy floor across the whole market, and several products in this comparison were built on it or on models trained in its wake. That is why the differentiation between commercial transcription tools in 2026 is almost never raw word accuracy. It is diarisation quality, streaming latency, language coverage, and whether the billing unit matches how you actually use the service.

Why diarisation being free matters

Speaker attribution is not a nice-to-have in most commercial speech workloads. Call analytics, meeting products, compliance recording and media subtitling all need to know who spoke. AssemblyAI charges for it, at $0.02 an hour for standard async diarisation and $0.12 an hour on streaming. Deepgram includes it for pre-recorded audio. On a workload of several thousand hours a month that difference compounds into a meaningful number, and it is the clearest reason to shortlist Deepgram for a build.

Model choice and language scope

Nova-3 ships in monolingual and multilingual variants at different prices, and Deepgram's model documentation lists coverage across English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch, Arabic and Chinese among others, with regional variants. Check the exact language list against your traffic before committing, because the multilingual premium is only worth paying if you actually receive that traffic.

Beyond a transcript

The reason to pick AssemblyAI over a cheaper engine is what comes back with the text. Entity detection, PII redaction, keyterm prompting and medical mode are all first-party, priced per hour and callable in the same request. For call-centre analytics, compliance monitoring or any pipeline that would otherwise chain a speech API to a separate language model, that consolidation removes a whole integration and its failure modes.

Costing a realistic configuration

Work the arithmetic before you compare vendors. Universal-3.5 Pro async at $0.21 an hour with standard diarisation at $0.02 and PII redaction at $0.08 is $0.31 an hour, which is materially above Deepgram's pre-recorded Nova-3 rate with diarisation already included. AssemblyAI earns that premium when you are buying the speech understanding features too, and does not when you only need text and speaker labels.

Streaming billing, stated plainly

AssemblyAI bills streaming on session duration, not audio duration, and says so on its pricing page. That is a design decision rather than a trap, but it changes how you architect a client. Close sockets aggressively on silence, or price the idle time into your unit economics.

Diarisation depth

Most engines degrade badly past five or six speakers. Scribe v2 documents support for up to 32, which is the difference between a usable transcript of a board meeting or a conference panel and a wall of unattributed text. Combined with precise word-level timestamps, it is also the most practical option here for generating subtitle files that have to carry speaker labels.

Language coverage, read properly

Scribe v2 supports more than 90 languages, and ElevenLabs groups them into published accuracy bands rather than claiming one number across all of them. English is in the top band at 5% word error rate or better. Languages in the lower bands can exceed 25%. Use the band your audio falls into as the planning number. A vendor that publishes the distribution instead of the best case is giving you something useful, so take it at face value and test accordingly.

Medical variant

Scribe v2 Medical is priced identically to the general model at $0.22 an hour. If your workload is clinical, that removes the usual premium other vendors attach to domain vocabulary. AssemblyAI, by contrast, prices medical mode as a $0.15 per hour add-on on top of the base model rate.

Live meeting integration

Otter joins calls as a participant, records, and produces a real-time transcript with live captions. Participants get a shareable link afterwards. Speaker identification works by matching voice patterns to participants and improves for recurring attendees. Accuracy degrades with crosstalk, poor microphones and heavily accented speech in the same way every engine on this page does, so the meeting hygiene matters more than the vendor choice.

Collaboration and search

Transcripts behave like documents. Team members highlight passages, comment and tag action items on the transcript itself, and search indexes every transcript so a decision can be found by keyword months later. For organisations with a recurring who-agreed-to-what problem, that searchable record is the actual product, and the transcription is the mechanism.

Where the caps bite

The two limits that generate complaints are the per-meeting length ceiling and the file import allowance. A 90 minute cap on Pro cuts off long workshops and all-hands calls. A ten-per-month import allowance rules out digesting an archive. Both are removed on Business, and neither is obvious from the plan headline, so check them against your real meeting pattern before buying seats.

Transcript-based editing

The editor presents a recording as a document. Selecting and deleting text removes the matching media. Reordering paragraphs reorders the timeline. Removing a tangent, resequencing a conversation or cutting a mistake becomes the same action as editing a doc. For anyone who has spent evenings trimming silences in a waveform editor, that is the whole pitch and it holds up.

Filler words, silence and Studio Sound

Filler word detection finds verbal fillers across the transcript and removes them in one pass, with the option to review each instance. Silence removal shortens long pauses on the same basis. Studio Sound applies enhancement that reduces background noise and evens out levels, which is the practical fix for remote recordings where participants had mismatched microphones. On interview-style content these three features are where the time saving actually comes from.

Two meters, not one

Descript bounds every plan by media hours and by monthly AI credits at the same time, and does not publish a per-action credit table. That means a plan can run out of either resource first depending on how heavily you lean on the AI features. Watch which meter you exhaust in the first month and size the plan from that, rather than from the media hours alone.

When a human is the requirement

Some contexts do not accept a probabilistic transcript. Legal transcription needs verbatim capture including false starts and crosstalk. Clinical records need correct terminology and correct attribution. Regulated recordings need a defensible record. In those settings the gap between a machine transcript and a reviewed one is not a quality preference, it is whether the artefact is usable at all, and $1.99 a minute is cheap against the cost of an error.

Reading the subscription against the a la carte price

Rev now sells AI transcription through seats rather than per minute, which makes cross-vendor comparison harder. Essentials includes 5,000 AI minutes per user per month for $25.49 per seat monthly. That is about 83 hours, which works out around $0.31 an hour if you use all of it, and considerably worse if you do not. The plans also carry discounts on human work, 10% on Essentials and 15% on Pro, which is the main reason to hold a seat at all.

Subtitles and localisation

Global subtitles are the line item that surprises buyers. Rev prices them per minute by target language, from $6.49 for Spanish and Hindi, through $10.49 for a group including French, German, Arabic and Portuguese, to $15.99 for Japanese and Korean. For a regular multilingual publishing schedule, price that against a machine translation pipeline with human review before committing.