Best speech to text models in 2026, and what the average error rate hides

For English audio, ElevenLabs Scribe v2 and Microsoft MAI-Transcribe-1.5 sit at the top of the hosted accuracy tables, at 2.2 and 2.4 percent word error rate. Among open weight models, Canary Qwen 2.5B leads on accuracy at 5.63 percent and Parakeet TDT 0.6B v3 leads on speed. Cost for 100 hours of audio runs from 2.52 dollars to 96 dollars.

Those four facts settle most shortlists. What they do not settle is which model handles your audio, and that is where the published averages stop being useful.

A speech to text model, also called an automatic speech recognition or ASR model, turns recorded or live audio into written text. The rankings below come from two public benchmarks, the model cards, and the pricing pages each vendor publishes.

What you need before you start

This piece assumes you are choosing a model, not building one. Before comparing anything, get four things in front of you.

• A sample of your real audio. Twenty to thirty minutes of the recordings you actually process. Sales calls, clinic dictation, podcast tape, support queues.

• A monthly volume figure. Hours per month, not minutes per day. Every price below is quoted per minute and every budget gets set per month.

• A latency requirement. Batch transcription that finishes in ten minutes is a different purchase from a live caption stream that has to return words in under 300 milliseconds.

• A deployment constraint. Whether audio may leave your network at all decides the open weight versus hosted question before accuracy does.

Who this fits: beginners picking a first transcription model, freelancers and small teams running under 50 hours a month, founders costing out a voice feature, and enterprise buyers who need language coverage and a license they can defend.

Who it does not fit: anyone looking for text to speech, voice cloning, or a finished meeting notes app. Those are separate categories with separate leaders.

How speech to text models get ranked

Three measurements decide almost every ranking you will see.

Word error rate, usually shortened to WER, counts how many words the model gets wrong per hundred spoken. Lower is better. A 5 percent WER means one word in twenty came out wrong, which sounds small until you notice it was the account number.

RTFx, the inverse real time factor, is total audio duration divided by transcription time. Parakeet TDT 0.6B v3 posts an RTFx of 3,332, meaning it processes roughly 3,332 seconds of audio per second of compute. Higher is better.

Latency matters only for streaming. It measures how long after a person speaks the first words appear.

Two public boards carry the numbers most articles quote. The Open ASR Leaderboard, run by Hugging Face with NVIDIA, Mistral AI and the University of Cambridge, evaluates more than 60 systems across 11 datasets. The Artificial Analysis speech to text leaderboard tracks hosted APIs on its own word error rate index alongside speed and price per 1,000 minutes.

Best Open source speech to text models

Open weight models ship their parameters under a license you can read, which means you can run them on your own hardware with no audio leaving your network.

Canary Qwen 2.5B pairs a speech encoder with a language model decoder, which is the architecture currently winning on English accuracy. It is slow relative to the field and English only.

Parakeet TDT 0.6B v3 is the throughput pick. NVIDIA's model card lists 600 million parameters, 25 European languages, and an average WER of 6.34 percent. At an RTFx above 3,000 it transcribes an hour of audio in roughly a second on server hardware.

Whisper large-v3 remains the default for anyone who needs breadth. OpenAI's repository puts it at 1,550 million parameters, about 10 GB of VRAM, and an MIT license. The turbo variant drops to 809 million parameters and 6 GB while running about eight times faster than large. Whisper reads audio through a sliding 30 second window, which shapes how it handles long recordings and silence.

Voxtral Mini Transcribe Realtime from Mistral is Apache 2.0, carries a 4B footprint sized for edge boxes, and lets you configure latency down to sub 200ms. The batch sibling accepts recordings up to three hours per request and takes up to 100 custom terms for context biasing.

Licenses split into two shapes here. MIT and Apache 2.0 let you ship without conditions. CC BY 4.0, which covers both NVIDIA models, requires attribution, and that requirement follows the model into your product.

Best hosted speech to text APIs

Hosted models trade control for accuracy you do not have to engineer.

Every row except Google comes from the Artificial Analysis word error rate index. Google publishes 0.016 per minute for standard recognition, which is 16.00 per 1,000 minutes.

ElevenLabs Scribe v2 holds the top accuracy slot on the Artificial Analysis index and the widest practical language list, at more than 90 languages. Its documentation sets file limits at 3 GB and 10 hours, dropping to 1 hour in multichannel mode.

Deepgram Nova-3 is the speed play, at a median speed factor of 541x, the fastest measured on that index.

AssemblyAI Universal is the pick when entity accuracy matters more than raw word error rate, since names, dollar figures, dates and product codes are what break downstream automation.

OpenAI gpt-4o-transcribe wins on integration rather than on any single number. The cheaper gpt-4o-mini-transcribe halves the rate.

Matching the model to the job

Four jobs pull in different directions.

Live voice agents. A voice agent has to return words before the caller notices a pause, so speed factor and streaming latency outrank accuracy. Deepgram Nova-3 and the realtime Voxtral build are sized for this.

Noisy environments. Call center floors, cars and shop floors degrade every model. The Open ASR Leaderboard includes the AMI meeting corpus for exactly this reason, so check per dataset scores rather than the headline average before buying for noisy audio.

Local and offline runs. Any model with published weights runs on your own machine with no network call. Parakeet and Whisper both have small variants that fit a laptop.

Multi speaker recordings. Diarization, which is the labeling of who spoke each line, is a separate feature. ElevenLabs, Voxtral and AssemblyAI ship it; several open models do not, and it bills as an add on where it exists.

What the average error rate hides

Every ranking above is an average across datasets. Change the dataset and the order changes with it.

The Open ASR Leaderboard splits its results into short form English, multilingual, and long form tracks precisely because models behave differently across them. On the long form track, Whisper large-v3 records a 6.43 percent WER at an RTFx of 68.56, while Parakeet CTC 1.1B records 6.68 percent at an RTFx of 2,793. Near identical accuracy, a 40x difference in throughput.

The pattern behind that gap is architectural. The leaderboard team reports that CTC and TDT decoders reach throughput 10 to 100 times higher than the alternatives, at slightly higher error rates.

The same effect shows in the opposite direction. Fine tuned Whisper variants beat the base model on English while losing multilingual coverage, so a model that improved on one board got worse on another.

Two practical consequences follow. A short form ranking says nothing about hour long recordings, and an English ranking says nothing about your Spanish queue. Two models a tenth of a percentage point apart are, for your purposes, the same model at different prices.

Audio length and file size limits, side by side

This is the constraint that breaks a working pipeline in week two.

The OpenAI guide caps uploads at 25 MB and accepts mp3, mp4, mpeg, mpga, m4a, wav and webm. A one hour stereo wav file is far past that ceiling, so anything longer than a short meeting needs compressing or splitting before it is sent.

Parakeet's ceiling is a memory ceiling rather than a policy one, which means it moves with your hardware.

What free actually gets you

"Free" covers two different things, and they fail in different ways.

Free as in open weights. Whisper, Parakeet, Canary Qwen, Qwen3-ASR and Voxtral cost nothing to download. You then pay for the GPU, the engineer who keeps it running, and the retry logic nobody budgets for.

Free as in a hosted allowance. AssemblyAI publishes up to 185 hours of pre recorded transcription and up to 333 hours of streaming at no cost and with no card. At its own pre recorded rate, that allowance is worth about 39 dollars. Deepgram offers 200 dollars of credit, also with no card. Google Cloud publishes no free monthly allowance on the v2 API at all, so billing starts at minute one.

For a first evaluation, the hosted free tiers are the faster route. A hundred and eighty five free hours covers a real pilot without provisioning anything.

What 100 hours of audio actually costs

I priced one workload from end to end: 100 hours of recorded calls a month, batch rather than streaming, English only, no add ons. Every figure below is a published per minute or per hour rate read from the vendor's own pricing page on August 18, 2026, multiplied out to 6,000 minutes.

Two rows come from the price column on the Artificial Analysis index rather than a vendor page, because neither ElevenLabs nor the cheapest batch provider publishes a per minute speech to text rate of its own.

The spread from cheapest to dearest is 38 times, on identical audio, for a difference in word error rate of a few percentage points. At 100 hours a month the whole decision is worth 93 dollars. At 10,000 hours it is worth 9,348, and the accuracy difference starts paying for itself in review time.

Two figures move that table more than the model choice does. Streaming costs more than batch at every vendor that prices them separately, roughly double at AssemblyAI. And add ons bill on top: speaker labels run 0.02 per hour there, medical mode 0.15.

Best speech to text models for Indian languages and Hindi

English rankings do not transfer. Hindi, Bengali, Tamil, Telugu and Marathi are underrepresented in the training data behind most models on this page, and the leaderboards that rank them mostly do not test Indic audio at all.

The purpose built option is IndicConformer 600M from AI4Bharat at IIT Madras. It covers all 22 officially recognized languages of India, ships under an MIT license, and offers both CTC and RNNT decoding. On the ARTPARK IISc Vaani benchmark it records a 13.2 word error rate for Hindi.

Read that next to the 2.2 percent that leads the English tables. Roughly one word in eight, against one word in fifty. The gap is the point: a system that is production ready in English is a review queue in Hindi.

Two general models are worth testing alongside it. Whisper large-v3 covers 99 languages including the major Indic ones. Meta's Omnilingual ASR reaches past 1,600 languages under Apache 2.0, which makes it the only realistic starting point for smaller regional languages.

Code switched audio, where a speaker moves between Hindi and English inside one sentence, degrades every model in this section. Test it directly rather than assuming a language flag handles it.

Hallucinated text on near silent audio

Speech to text models built on language model decoders can output fluent sentences that were never spoken. It happens most on silence, background noise, and disfluent speech.

The size of the problem is measured. A study presented at the ACM Conference on Fairness, Accountability, and Transparency analyzed 13,140 audio segments from AphasiaBank and found that "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences". Of those hallucinations, 38 percent included explicit harms such as invented violence or false attributions of authority.

The rate was higher for speakers with longer non vocal pauses, a common symptom of aphasia. Any audio with long gaps carries the same exposure: legal recordings, medical dictation, interview tape, hold music on a support line.

Two defenses work in practice. Strip silence with a voice activity detector before transcription, which removes the input that triggers most of it. And check whether the model returns per word confidence scores, since a hallucinated span usually scores low even when it reads perfectly.

The 30 minute check on your own audio

No leaderboard tests your recordings. This check runs on the free allowances and produces a number you can defend.

Stage one, build the reference. What happens: take 10 minutes of your real audio and transcribe it by hand, or correct a machine transcript until it is exact. Include the accents, the crosstalk and the jargon you actually get. Done when: you have a verbatim text file you would sign off on.

Stage two, run three candidates. What happens: send the same file to one hosted API, one open weight model, and one cheaper tier of the hosted API. Save each output unedited. Done when: three transcripts exist for one audio file.

Stage three, count the errors that matter. What happens: count total wrong words for the WER figure, then count separately the errors on names, numbers, dates and product terms. Done when: you have two error counts per model rather than one.

Stage four, decide on the second number. What happens: compare the entity error count for each model against its cost per 100 hours of audio. Done when: you can name the model you are choosing and the number that chose it.

Copy this in as your scoring sheet:

AUDIO SAMPLE: ____ minutes, ____ speakers, ____ language(s)
MONTHLY VOLUME: ____ hours

MODEL 1: ______________
  Total word errors: ____ / ____ words = ____% WER
  Entity errors (names, numbers, dates, terms): ____
  Cost per 100 hours: $____
  Longest single file accepted: ____
  Notes: ______________

MODEL 2: ______________
  [repeat]

MODEL 3: ______________
  [repeat]

DECISION: ______________
REASON (one number): ______________

Four mistakes that cost people a rewrite

Ranking on the wrong track. A model chosen from a short form English table then deployed on 45 minute recordings is a different model in production. Check the long form numbers before committing.

Treating word error rate as one number. A transcript can score 4 percent WER and still get every account number wrong. Filler words are cheap to miss. Entities are not.

Ignoring the license until launch. CC BY 4.0 on the NVIDIA models requires attribution downstream. MIT and Apache 2.0 do not. This is a five minute check that becomes a legal review if it is skipped.

Assuming faster hardware fixes accuracy. It does not. Practitioners in a Hacker News thread on running Parakeet locally, which drew 114 points and 31 comments, reported inference under 100 milliseconds for 60 seconds of audio on a laptop, while the same discussion kept circling back to language coverage as the real constraint. Speed and accuracy are separate purchases.

Frequently asked questions

What are the best speech to text models right now? For English accuracy, ElevenLabs Scribe v2 at 2.2 percent word error rate leads the Artificial Analysis index. For open weights, Canary Qwen 2.5B leads accuracy at 5.63 percent and Parakeet TDT 0.6B v3 leads speed at an RTFx above 3,000.

What is the best free speech to text model? Whisper large-v3 under an MIT license is the most flexible free option, at 99 languages and no usage conditions. If free means a hosted allowance instead, AssemblyAI publishes up to 185 free hours of pre recorded transcription.

Which speech to text model is best for Hindi and Indian languages? IndicConformer 600M from AI4Bharat covers all 22 scheduled languages of India under an MIT license and records 13.2 WER for Hindi on the Vaani benchmark. Whisper large-v3 is the general purpose fallback.

Is Whisper still the best open source speech to text model? Not on accuracy or speed. Canary Qwen 2.5B and Parakeet TDT 0.6B v3 both beat it on the Open ASR Leaderboard. Whisper still wins on language breadth at 99 languages and on the size of its tooling ecosystem.

Where is the speech to text leaderboard? Two are worth reading. The Open ASR Leaderboard on Hugging Face covers open and proprietary systems across 11 datasets. Artificial Analysis tracks hosted APIs on word error rate, speed factor and price.

Which speech to text models are on Hugging Face? Whisper, Parakeet TDT, Canary Qwen, Qwen3-ASR, Voxtral, Moonshine, wav2vec 2.0 and IndicConformer all publish weights and model cards there, most with a runnable demo attached.

What to do with this next

Pick two models rather than one: the accuracy leader in your language, and the cheapest option that clears your latency requirement. Run both through the 10 minute sample, then compare entity errors before total word errors. The published averages narrow the field to a shortlist. Your own audio picks the winner, and it takes half an hour to find out.

Sources, model cards and vendor pricing rechecked on August 18, 2026.