So choosing an automatic speech recognition engine on headline claims doesn't work, because every vendor publishes its own best number on its own test set. This review puts 11 managed ASR engines in one format, with the current model, accuracy from a single independent benchmark, language coverage, real-time support, diarization, and price, so you can build a shortlist and then test it on your own audio.
TL;DR
- ASR engines now differ less on clean English and more on noisy conversational speech, language coverage, latency, and price per audio hour.
- This page compares 11 managed engines in the same format, using vendor documentation and the Artificial Analysis word error rate index, where scores range from 2.0% to 5.2% across the engines listed.
- List prices for each engine's current batch model run from about $0.10 to $0.38 per audio hour, and one of the two lowest is a promotional preview rate.
- There's no pick and no ranking: shortlist by use case, then test the shortlist on your own audio.
How we compared ASR engines
Every vendor publishes accuracy numbers, but each one picks its own test set, so their figures rarely line up. To keep this comparison fair, we held every engine to the same seven criteria:
- Current model
- Published accuracy
- Languages
- Real-time support
- Speaker diarization
- Deployment options
- List price per audio hour for batch transcription, taken from each vendor’s own pricing page
Every figure was checked against a live source in October, 2026. For accuracy, we used one independent benchmark for all 11 engines, so the numbers can sit side by side: the Artificial Analysis speech-to-text leaderboard. Its AA-WER index averages word error rate (WER, the share of words a model gets wrong against a reference transcript) across about 8 hours of audio from three datasets, weighted by duration. The datasets are speech directed at voice agents, European Parliament recordings and earnings calls. For open-source models, we also refer to the Hugging Face Open ASR Leaderboard. Features, languages and launch dates come from vendor documentation and independent coverage, linked in each section.
A note on bias: Gladia makes one of the engines listed. It gets the same format, sources and scrutiny as every other provider, and where other engines score better, we say so.
Two caveats before you read the numbers. First, AA-WER doesn’t break results down by language, so it tells you little about accuracy in languages like Polish, Hindi or Arabic. Second, prices move. Microsoft’s rates are promotional until December 31, 2026, Deepgram’s streaming rates are limited-time offers, and several models are still in preview. The table marks ea
What an ASR engine does: from speech signal to written text
When people compare speech-to-text providers, they usually compare models. What you’re really choosing is an ASR engine: the full system that converts spoken language into written text. It takes in the audio, runs the ASR model, and then uses post-processing to turn raw word sequences into a transcript someone can work with.
That last part is easy to underestimate. The model decides which words it heard. The engine decides whether you get accurate transcripts on noisy phone calls, timestamps you can sync subtitles to, and speaker labels when a meeting has multiple speakers. That’s why this review compares engines, not just models. It also helps to separate speech recognition from voice recognition, because the two terms often get mixed up. Voice recognition identifies the speaker rather than the words spoken. Voice recognition software answers “who is talking?”, while an ASR engine answers “what did they say?”
Even within speech recognition, the same automatic speech recognition technology does two quite different jobs. One is handling short voice commands for assistants. The other is automatic transcription of recorded audio, video files and live streams. This review is about the second: engines that turn human speech in calls, meetings and media into text you can search and analyze.
From Bell Labs to deep learning models
Every generation of speech recognition has removed a limit the previous one couldn’t get past.
The first limit was the speaker. In 1952, Bell Labs built Audrey for single speaker digit recognition: it recognized the digits zero to nine, spoken by one person. Because human voices vary so much, Audrey’s accuracy improved the more it worked with a given speaker, so every new voice was a new problem.
The next step swapped exact matching for probability. James Baker’s 1975 dissertation applied hidden Markov models to work out which words most likely produced a given sequence of sounds, work that helped develop the first commercially viable speech recognition software. Statistical models of this kind, paired with separate acoustic modeling and a language model, ran most ASR systems until deep neural networks overtook them. The catch was the pipeline: such systems needed each stage tuned by hand, which made every new language slow and expensive to add.
Deep learning, the branch of machine learning built on large neural networks, removed that bottleneck. Instead of experts tuning each stage, the model learns the whole mapping from audio data. OpenAI’s Whisper, published in December 2022, made the shift obvious: it was trained on 680,000 hours of weakly supervised multilingual audio as a single model that maps audio straight to text across many languages. Most modern ASR engines now follow that end-to-end approach in some form. That’s why the differences between engines today come less from the architecture and more from the training data and everything built around the model.
Inside a modern ASR engine
A modern ASR engine still runs a recognizable pipeline, even when one neural network does most of the work:
- Audio handling: The engine decodes the file or stream, resamples it, and may apply noise reduction before recognition.
- Feature extraction: The engine cuts audio signals into short frames and turns them into a spectrogram-like representation the model can read.
- Recognition: The ASR model predicts word sequences from those features. In older systems, this was acoustic modeling plus a separate language model; in end-to-end systems, a single model does both.
- Post-processing: The engine adds punctuation, casing, number formatting, word timestamps, speaker labels, and sometimes translation to produce the final output. Some engines then run natural language processing or other AI models on the transcript for summaries or sentiment.
Two engines can share the same class of model and still produce very different transcripts, because training data, post-processing, and features such as custom vocabulary, used to improve accuracy on domain-specific terminology, all change the result. That's why the comparison below covers the engine, not only the model.
The engines, one by one
Each engine below follows the same pattern: the current model, how it works as far as the vendor has published, strengths, trade-offs, pricing, and where it fits best. Engines are in alphabetical order.
Speech-to-Text Provider Comparison
| Provider |
Current model(s) |
AA-WER (batch) |
Languages |
Real-time |
Diarization |
Deployment |
List price per audio hour (batch) |
| Amazon Transcribe |
Speech foundation model (no public name) |
4.1% |
100+ |
Yes |
Yes, included |
AWS cloud |
$0.36 |
| AssemblyAI |
Universal-3.5 Pro, Universal-3.6 Pro Realtime, Universal-2 |
3.1% (Universal-3 Pro) |
18 on 3.5 Pro, 32 on 3.6 Pro Realtime, 99 with Universal-2 |
Yes |
Yes, add-on |
Cloud API, US and EU regions |
$0.21 |
| Deepgram |
Nova-3, Flux, Flux Multilingual |
5.2% (Nova-3) |
45+ across Nova models, 10 in multilingual mode |
Yes |
Yes, included in batch |
Cloud API (US and EU endpoints), self-hosted |
$0.26 |
| ElevenLabs |
Scribe v2, Scribe v2 Realtime, Scribe v2 Medical |
2.2% (Scribe v2) |
90+ |
Yes |
Batch only |
Cloud API |
$0.22 |
| Gladia |
Solaria-1, Solaria-3 |
4.1% (Solaria-1), 3.2% (Solaria-3) |
100+ on Solaria-1, 5 on Solaria-3 |
Yes (Solaria-1) |
Async only, included |
Cloud (separate EU and US infrastructure), on-premise |
€0.27 (Solaria-1), €0.16 (Solaria-3) |
| Google Cloud |
Gemini 3.5 Transcribe, Chirp 3 |
2.6% (Gemini 3.5 Transcribe) |
85+ |
Yes |
Yes |
Global endpoint for now |
About $0.30 (token-based estimate) |
| Microsoft |
MAI-Transcribe-2, MAI-Transcribe-2-Streaming (both preview) |
2.0% |
60 |
Yes (separate streaming model) |
Batch only |
Four Azure regions during preview |
$0.10 (promotional to Dec 31, 2026) |
| Mistral |
Voxtral Mini Transcribe V2, Voxtral Realtime |
3.6% (Mini Transcribe V2) |
13 |
Yes |
Batch only |
Cloud API, open weights for Realtime |
$0.18 |
| OpenAI |
GPT-Transcribe, GPT-Live-Transcribe |
3.3% (GPT-Transcribe) |
Multilingual, with language hints |
Yes |
Separate, deprecated model |
Cloud API |
$0.27 |
| Soniox |
stt-async-v5, stt-rt-v5 |
3.8% (v5 async) |
60+ |
Yes |
Yes, included |
Cloud API, US, EU and Japan regions |
$0.10 |
| Speechmatics |
Melia 1, Enhanced, Standard, Oak 1, Linden 1 |
4.9% (Melia), 4.0% (Enhanced) |
55+ |
Yes (Enhanced, Standard, Linden 1) |
Yes, included |
Cloud, on-premise, on-device |
$0.12 (Melia 1), $0.38 (Enhanced) |
Amazon Transcribe
If your audio already lives in AWS, the shortest path to a transcript is Amazon Transcribe. It runs on a speech foundation model with no public model name, through batch and streaming APIs, with separate versions built for medical and contact center audio.
How it works: Amazon describes a multi-billion-parameter model trained on millions of hours of speech data with self-supervised learning, launched in November 2023 to take coverage past 100 languages. Training data was sampled to balance languages, so underrepresented languages carry more weight than their raw share of the audio. At launch, Amazon reported accuracy gains of 20% to 50% for most languages and 30% to 70% on telephony speech.
Strengths: The appeal is breadth rather than one headline number. You get 100+ languages for batch, a long-running streaming API over HTTP/2 and WebSockets, and speaker diarization, custom vocabularies and language identification included in the base rate. Transcribe Medical and Call Analytics, which added generative call summaries with the foundation model release, cover two verticals without a second vendor.
Trade-offs: The accuracy gains Amazon announced applied to batch mode. On Artificial Analysis, Transcribe scores 4.1% AA-WER, mid-table here. It processes about 30 seconds of audio per second, which puts it among the slower batch engines in this comparison. PII redaction and custom language models cost extra.
Pricing: In US East, $0.006 per minute for batch ($0.36 per hour) and $0.01per minute for streaming ($0.60 per hour). Rates vary by region.
Best for: Teams already on AWS that need call analytics or medical transcription inside the same cloud.
AssemblyAI
Two models do most of the work at AssemblyAI. Universal-3.5 Pro handles recorded audio and was released in July 2026. Universal-3.6 Pro Realtime handles streaming and was released on September 29, 2026. Universal-2 sits behind both to cover more languages.
How it works: Universal-3.5 Pro takes contextual prompts and keyterms to steer recognition toward the words that matter in your audio. Universal-3.6 Pro Realtime is trained on voice-agent and telephony conversations. It detects the language automatically and keeps the turn open while a caller reads out a number or an email address. When a request falls outside a model’s languages, AssemblyAI falls back to Universal-2.
Strengths: Native code switching covers 18 languages on Universal-3.5 Pro and 32 on Universal-3.6 Pro Realtime, with 99 through the Universal-2 fallback. Artificial Analysis lists Universal-3 Pro at 3.1% AA-WER and about 109 seconds of audio processed per second. Universal-3.5 Pro also powers two newer APIs: a Sync API that returns a transcript in a single request, and a Dictation API that removes filler words and resolves self-corrections. The EU region costs the same as the US.
Trade-offs: The catch is the language list. Flagship batch accuracy covers 18 languages, and anything else runs on the older Universal-2. Add-ons also stack on the base rate: diarization is +$0.02 per hour in batch and +$0.12 in streaming, and key terms and prompting are +$0.05 each in batch.
Pricing: $0.21 per hour for Universal-3.5 Pro, $0.15 for Universal-2, and $0.45 for Universal-3.6 Pro Realtime.
Best for: Products in English and other major languages that want prompting and one vendor for batch, streaming, and dictation.
Deepgram
If you're looking for a leader in raw speed, Deepgram is a solid one. Its lineup has two parts: Nova-3 for batch and streaming, and Flux for conversational streaming with end-of-turn detection built in.
How it works: Flux combines recognition with model-native turn detection, so the model that transcribes also signals when a speaker has finished. Deepgram documents end-of-turn detection at about 260 ms, with a threshold you can raise for fewer false turns or lower for faster ones. Flux comes as an English model and a multilingual model that covers 10 languages, with optional language hints.
Strengths: Artificial Analysis measures Nova-3 at about 601 seconds of audio processed per second, almost twice the next engine here. Deepgram’s Nova models cover 45+ languages, Nova-3 comes in medical and pharmaceutical variants, and diarization is included in batch. For teams with data constraints, Deepgram offers an EU endpoint and self-hosted deployment.
Trade-offs: Speed comes with an accuracy cost on this benchmark: Nova-3’s 5.2% AA-WER is the highest batch score in this table. Its multilingual mode with code-switching covers 10 languages, and other languages run on monolingual models. Streaming rates are also promotional; the regular Nova-3 and Flux English price is $0.0077 per minute.
Pricing: $0.0043 per minute (0.26 per hour) for Nova-3 batch, $0.0048 per minute promotional for Nova-3 streaming and $0.0065 for Flux English. Streaming diarization adds $0.0020 per minute.
Best for: Real-time voice agents where latency and turn-taking matter as much as raw accuracy.
ElevenLabs
The second-most accurate engine on Artificial Analysis’s index comes from ElevenLabs: Scribe v2, launched in January 2026. It sits alongside Scribe v2 Realtime for live transcription and Scribe v2 Medical for clinical audio.
How it works: ElevenLabs hasn’t published architecture details for Scribe v2. Scribe v2 Realtime is a separate streaming model that returns partial transcripts in about 150 ms, according to ElevenLabs.
Strengths: At 2.2% AA-WER, Scribe v2 trails only Microsoft in this table. It covers 90+ languages and accepts up to 1,000 keyterms. It also supports diarization for up to 32 speakers, audio-event tagging such as laughter or music, and entity detection. ElevenLabs lists Scribe v2 Medical as HIPAA-eligible.
Trade-offs: Those strengths are mostly batch strengths. The Realtime model doesn’t label speakers, and it costs almost twice as much as batch. Keyterm prompting and transcript editing are billed per hour on top of the base rate.
Pricing: $0.22 per hour for Scribe v2 and Scribe v2 Medical, with keyterm prompting at +$0.05 and transcript editing at +$0.07. Scribe v2 Realtime is $0.39 per hour.
Best for: Multilingual media and meeting transcription where batch accuracy matters more than live speaker labels.
Gladia
Built around two in-house models, Gladia splits the work by audio type. Solaria-1 covers 100+ languages, code-switching, and real-time. Solaria-3 targets noisy European business audio in English, French, German, Spanish, and Italian.
How it works: Both are Gladia’s own models, served through one API, and you pick the model per request. Solaria-1 detects the language as it goes, so when a speaker switches mid-sentence, the transcript follows. Solaria-3 detects which of its five languages it hears and holds that language for the whole file.
Strengths: On Gladia’s open benchmark, which sends identical files to each provider’s production API and is open-sourced, Gladia ranks first across:
- Switchboard conversational phone audio: Solaria-3 scores 33.9% WER, 8.4 points ahead of the next competitor. It beats AssemblyAI, Speechmatics, Mistral Voxtral, Deepgram Nova-3 and ElevenLabs Scribe v2.
- DIHARD III diarization: Solaria-1 scores a 16.6% weighted diarization error rate. Here it beats NVIDIA NeMo Sortformer, pyannoteAI Community-1, Speechmatics, AWS Transcribe, Soniox, ElevenLabs Scribe v1, OpenAI GPT-4o Transcribe, AssemblyAI Universal and Deepgram v3.
These are Gladia’s own measurements. On the independent Artificial Analysis index, Solaria-3 scores 3.2% AA-WER, ahead of OpenAI, Mistral, Soniox, Speechmatics, Amazon and Deepgram. Diarization, custom vocabulary, and name consistency are included. Gladia is GDPR and HIPAA compliant, SOC 2 Type 2 certified, and keeps EU and US workloads on separate infrastructure.
Trade-offs: Currently, diarization is async only. For live audio, the workaround is to send each speaker on a separate channel: Gladia then tags every utterance with its channel, so you get speaker-attributed transcripts in real time without diarization.
Pricing: Async starts at €0.27 per hour on Solaria-1 and €0.16 on Solaria-3, and the rate drops automatically as your monthly volume grows, down to €0.21 and €0.12. Larger top-ups also earn up to 20% in bonus credits, and new accounts start with €50 free. Real-time is €0.68 per hour. Translation, named entity recognition, and PII redaction are optional add-ons, billed only when you use them. See the full pricing for every tier.
Best for: Multilingual meeting assistants and European contact centers handling noisy, accented, or code-switched audio.
Google Cloud
Google now has two speech-to-text paths, and the newer one comes from the Gemini side of the company. Gemini 3.5 Transcribe, announced on August 26, 2026, ships with a separate live model for streaming. Chirp 3 remains available in Speech-to-Text V2.
How it works: Gemini 3.5 Transcribe builds on Gemini’s audio and language understanding and can return cleaned-up text. It can remove filler words and disfluencies, handle self-corrections and format output automatically. Google’s pricing page lists speaker diarization, word-level timestamps and custom vocabulary biasing.
Strengths: Gemini 3.5 Transcribe detects languages automatically across 85+ languages, and Google puts its time to final transcription at 70% faster than Chirp 3. Independent numbers support the upgrade: 2.6% AA-WER, among the lowest in the table, and about 91 seconds of audio per second against Chirp 3’s 29.
Trade-offs: Google’s own sources disagree on its status. The launch post calls it a public preview, while the Gemini API docs list both models as generally available. For now it runs on a global endpoint, with single-region support listed as coming. Cleaned-up output also isn’t a verbatim transcript, which matters for compliance and legal records.
Pricing: Gemini 3.5 Transcribe is billed by tokens: Google estimates about $0.005 per minute (about $0.30 per hour), and about $0.009 per minute for the live model. Chirp 3 is $0.016 per minute on Speech-to-Text V2 standard recognition, or $0.003 with 24-hour dynamic batch.
Best for: Teams already in Google Cloud, or building on Gemini, that want readable transcripts rather than verbatim ones.
Microsoft Azure
The lowest error rate in this comparison belongs to a model that’s still in preview. MAI-Transcribe-2, Microsoft’s batch engine, was released in public preview on September 3, 2026. Its first streaming model, MAI-Transcribe-2-Streaming, followed on October 1.
How it works: Both are Microsoft AI’s in-house models, served through Azure Speech and Microsoft Foundry. Batch output comes in a clean or verbatim style, so you choose between a readable transcript and a word-for-word one.
Strengths: At 2.0% AA-WER, MAI-Transcribe-2 is the most accurate engine here. It’s also the fastest after Deepgram on Artificial Analysis, at about 318 seconds of audio per second. It covers 60 languages, up from 25 on MAI-Transcribe-1, with an average 5.2% WER across them on the FLEURS benchmark. Diarization, word-level timestamps and keyword biasing come at no extra charge in batch.
Trade-offs: That preview label matters. Both models are public previews with no service-level agreement. Microsoft’s own comparison lists no diarization, word timestamps or keyword biasing for the streaming model. During preview, batch runs in four regions: East US, West US, Southeast Asia and North Europe.
Pricing: $0.10 per hour for batch and $0.54 per hour for streaming. Azure’s pricing page labels the batch rate a limited-time offer through December 31, 2026 that reflects preview pricing. Microsoft calls the streaming rate introductory through the same date. No standard rate has been published.
Best for: Enterprises on Azure that can run a preview model and budget for a price change in 2027.
Mistral
For teams that want to run real-time transcription on their own hardware, Mistral is the only vendor here that publishes open weights for it. Its speech family, released in February 2026, pairs Voxtral Mini Transcribe V2 for batch with Voxtral Realtime for live audio.
How it works: Voxtral Realtime is a 4B-parameter streaming model with open weights under Apache 2.0. It runs on a single GPU with at least 16GB of memory and is supported in vLLM. Its transcription delay is configurable from 80 to 2,400 ms, with 480 ms recommended.
Strengths: On the batch side, Voxtral Mini Transcribe V2 includes diarization, context biasing and word timestamps at 3.6% AA-WER. Mistral’s larger Voxtral Small is the most accurate open-weights model on Artificial Analysis, at 2.8%.
Trade-offs: Language coverage is the main limit. Both models cover 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian and Dutch. Diarization is on the batch model only, so a live product that needs speaker labels has to add them elsewhere.
Pricing: $0.003 per minute(0.18 per hour) for batch and $0.006 per minute ($0.36 per hour) for the hosted Realtime API.
Best for: Cost-sensitive batch work in major languages, and teams that want a real-time model they can host themselves.
OpenAI
Between July and August 2026, OpenAI replaced its transcription lineup. GPT-Transcribe for recorded audio and GPT-Live-Transcribe for live streams arrived on July 28, and the older models were deprecated a month later.
How it works: Both models accept free-form context, keyword hints, and multiple language hints to steer recognition toward specialized terms and multilingual speech. GPT-Live-Transcribe adds tunable latency for streaming. On OpenAI’s own launch benchmark across 22 Common Voice languages, GPT-Transcribe roughly halved Whisper-1’s error rate; that’s OpenAI’s measurement, not an independent one.
Strengths: Simplicity is the draw. GPT-Transcribe scores 3.3% AA-WER at $0.0045 per minute, 25% less than Whisper-1. If you already build on OpenAI, it’s one more call against the same account and billing.
Trade-offs: At launch, neither model returned speaker diarization, word-level timestamps, SRT or VTT subtitles, or English translation. The older models that filled those gaps, including whisper-1 and gpt-4o-transcribe-diarize, were deprecated on August 26, 2026, and will leave the API on February 26, 2027.
Pricing: $0.0045 per minute ($0.27 per hour) for GPT-Transcribe and $0.017 per minute (about $1.02 per hour) for GPT-Live-Transcribe, nearly four times the batch rate.
Best for: Teams already on OpenAI that need transcripts without speaker labels.
Soniox
Where several engines switch models depending on the language, Soniox runs a single model for all of them. The current versions are stt-async-v5 for recorded audio, released in June 2026, and stt-rt-v5 for real time.
How it works: With one model across every supported language, a mid-sentence language switch doesn’t need a second model or a restart. Soniox says v5’s largest gains over v4 are on noisy recordings, phone calls, far-field microphones and multi-speaker meetings.
Strengths: Soniox handles 60+ languages with mixed-language audio in real time, with a median time to final segment of about 250 ms in Pipecat’s default setup. Diarization, language identification and formatting are included in the base rate, and translation runs in the same stream, one-way or two-way. The async model scores 3.8% AA-WER, and Soniox offers US, EU, and Japan regions.
Trade-offs: The single-model design has a side effect: language hints bias the model without blocking other languages. If your audio must stay in one language, test how often it drifts. Async processing is also slower than most engines here, at about 43 seconds of audio per second on Artificial Analysis, which matters only for large backlogs.
Pricing: About $0.10 per hour async and $0.12 real-time. With translation, the rate is about $0.16 async and $0.18 real-time.
Best for: Multilingual real-time apps and live translation.
Speechmatics
Since June 2026, Speechmatics has added three models to its lineup. Melia 1 handles multilingual audio, Linden 1 powers its Agent STT for voice agents, and Oak 1 covers medical audio. Enhanced and Standard remain its established models.
How it works: Melia 1 adds code-switching across all 55+ supported languages in a single pass, with no language to select in advance. Oak 1, released on October 14, 2026, applies the same multilingual approach to clinical conversations.
Strengths: Deployment range sets Speechmatics apart: SaaS, on-premise containers, and on-device, a combination few engines here match. Diarization, word timings, and smart formatting are included in every model. Enhanced scores 4.0% AA-WER, and Melia processes about 166 seconds of audio per second on Artificial Analysis.
Trade-offs: The newest models aren’t yet the most complete. Melia 1 and Oak 1 are batch only for now, so live code-switching still needs another engine. Melia’s 4.9% AA-WER is higher than Enhanced’s 4.0%, so the multilingual model costs some accuracy on Artificial Analysis’s index. Custom dictionary is available on Enhanced and Standard but not on Melia.
Pricing: $0.12 per hour for Melia 1, $0.15 for Oak 1, $0.23 for Standard, and $0.38 for Enhanced, with a 33% discount if you allow model training on your audio.
Best for: Regulated industries and broadcasters that need on-premise or on-device deployment.
Open-source ASR models you can host yourself
Open-source ASR software removes the per-hour fee and moves every other cost to your team: GPUs, scaling, diarization, monitoring, and model updates. It makes sense when your data can’t leave your own servers, or when you transcribe so much audio that running it yourself costs less than paying per hour. Four families lead today:
- Whisper large-v3 and large-v3 turbo from OpenAI, the reference open model since its 2022 release. Hosted versions of Whisper Large v3 score from 4.1% to 10.1% AA-WER depending on the provider serving them, which shows how much the engine around a model matters.
- NVIDIA Parakeet and Canary, with Parakeet TDT 0.6B V3 at 4.5% and Canary Qwen 2.5B at 4.3% AA-WER on Artificial Analysis.
- Qwen3-ASR from Alibaba's Qwen team, covering 30 languages and 22 Chinese dialects.
- Meta Omnilingual ASR, released in November 2025, which extends recognition to more than 1,600 languages under an Apache 2.0 license.
Our guide to the best open-source speech-to-text models covers each in more depth, including hardware needs.
How to choose an ASR engine for your use case
The right engine depends on the audio you send and what you do with the transcript afterwards. Start from your use case, list the two or three criteria that would break your product if they failed, and drop every engine that misses one.
Note-takers and meeting assistants
Meeting assistants live or die on who said what. Diarization quality, measured as diarization error rate (DER), matters more than a fraction of a point of WER, because a summary that assigns the decision to the wrong person is worse than a typo. Check long-file accuracy on recordings of an hour or more, word timestamps for search and playback, and code-switching if your users mix languages. Several engines on this list offer diarization on batch only, which is fine for notes produced after the call.
Contact centers
Contact center audio is the hardest test for most ASR technology: narrowband phone calls with crosstalk, where accurate transcription of a fast-spoken account number decides whether the record is usable. Test accuracy on real 8kHz recordings, since telephony audio breaks many general-purpose models. Then check PII redaction, data residency for your regulators, and batch price at your monthly volume, because QA and analytics run on recorded audio.
Voice agents
Now, a voice agent needs a final transcript fast enough to answer before the caller notices a gap. Compare streaming latency, measured as time to final transcript, along with how each provider measures it. End-of-turn detection, real-time accuracy, and concurrency limits per account matter as much as WER. Speaker diarization rarely matters here, since there's usually one caller.
Test on your own audio before you commit
Published benchmarks narrow the field; your own audio and your own data seal the deal. Take 50 to 100 files that represent your real traffic, including the worst ones with the most background noise and accents, and score each shortlisted engine on WER, DER, and latency. As a fun bonus, our blind comparison tool lets you upload audio and vote on transcripts side-by-side with model names hidden, which could be a cool gut check before you run your own official benchmark.
To include Gladia in your test, sign up for the playground and run your own files through Solaria-1 and Solaria-3 with €50 of free credit.
FAQs
What is an ASR engine?
An ASR engine is the complete system that converts spoken audio into written text: the ASR model plus audio handling, language detection and post-processing such as punctuation, timestamps and speaker labels. Gladia’s API, for example, wraps its Solaria models with language detection, custom vocabulary and speaker diarization in a single request. Our guide to automatic speech recognition explains how the models inside it work.
What are the best ASR engines?
The main managed ASR engines in 2026 are Gladia, Amazon Transcribe, AssemblyAI, Deepgram, ElevenLabs, Google Cloud, Microsoft Azure, Mistral, OpenAI, Soniox and Speechmatics. None is best for every case, so shortlist by languages, audio conditions, and latency needs, then test on your own audio. If your audio is noisy, accented, or switches language mid-sentence, Gladia’s Solaria models are built for exactly that.
What's the difference between an ASR engine and an ASR model?
The model is the trained neural network that predicts text from audio, such as Whisper or Solaria-1. The engine is the full system around it, including pre-processing and the post-processing that adds punctuation, timestamps, diarization and formatting. That’s why two engines built on similar models can still differ widely in accuracy and features.
Which ASR engine is most accurate for multilingual audio?
It depends on the language and the audio, so no single engine wins everywhere. Independent indexes such as Artificial Analysis don’t break results down by language, so ask vendors for per-language word error rate and test code switching on your own recordings. Gladia publishes its multilingual results, including where it trails, on its open benchmark.
How much does an ASR API cost per hour?
List prices for batch transcription run from about $0.10 to $0.38 per audio hour across the engines in this review, as of October 16, 2026, while real-time and add-ons usually cost more. Most providers lower the rate as volume grows. Gladia’s async rates, for example, start at €0.16 per hour on Solaria-3 and drop automatically as your monthly volume rises.
How does automatic speech recognition work?
Automatic speech recognition turns an audio signal into features, uses a neural network to predict the most likely word sequences, then formats the result into readable text. Older systems split this into acoustic modeling and a separate language model, while modern end-to-end models such as Whisper do both in one network. Our explainer on how speech recognition models work goes step by step.
What does ASR stand for?
ASR stands for automatic speech recognition, the technology that converts speech to text without a human transcriber. The term covers both the models and the engines and APIs built around them.
Glossary
WER: Word error rate: the share of words a model gets wrong against a reference transcript, counting substitutions, deletions and insertions. Lower is better.
DER: Diarization error rate: the share of audio time with a speaker error, such as missed speech or speech attributed to the wrong speaker.
Diarization: Labeling which speaker said each part of a recording. See speaker diarization.
Code-switching: A speaker changing language within a conversation or a sentence. See what code-switching is.
Hallucination: Text an ASR model outputs that was never spoken, often during silence or noise.
Phoneme: The smallest unit of sound that distinguishes one word from another.
Spectrogram: A representation of an audio signal's frequencies over time, the input most ASR models read.
Real-time factor (RTF): Processing time divided by audio duration; below 1 means faster than real time. See how latency is measured.
Streaming vs batch: Streaming transcribes audio as it arrives, for live use; batch transcribes a complete file. See real-time and async transcription.
