November 05, 2026
Whisper vs Deepgram vs AssemblyAI: dictation accuracy 2026
Whisper vs Deepgram vs AssemblyAI has no defensible accuracy winner for every dictation job. Test exact models on the same recordings, using the same language and output rules. Then run a separate trial with vocabulary hints. Otherwise, you may be comparing configuration choices rather than recognition quality.

This guide helps developers choose a speech engine for short, deliberate dictation. It isn’t a voice-agent leaderboard or a report on tests we ran. DictaFlow publishes this guide as a dictation-app vendor. The examples below are proposed evaluation materials, not measured results.
Name the implementation before choosing a winner
Whisper is an open-source model family, not one fixed service. OpenAI's repository documents several model sizes, English-only variants, and multilingual models. Its published speed comparisons use specific hardware. A locally hosted model needs a separate configuration record from a managed endpoint, even when both are described as Whisper.
For Deepgram and AssemblyAI, record the requested model, endpoint, and request date. Save the model version each service returns, when available. First, compare file transcription with file transcription. Streaming adds decisions about when speech ends and when text becomes final.
Keep a short configuration sheet with the model ID, language, audio format, vocabulary hints, punctuation settings, and any later rewriting. Save it alongside the returned text. The broader dictation benchmark guide covers application-level buying criteria; this comparison focuses only on choosing an engine.
Keep development examples out of the final test
Create two small sets using the kind of speech your application will receive. Use one to choose settings, and leave the other untouched until you’ve settled on them. This keeps you from tuning every option to fit the examples you later present as evidence.
For example, imagine a fictional publishing setup with an editor named Niamh, an imprint called Alder Finch, and the term EPUB. Include ordinary sentences along with difficult terms. A recognizer hasn’t solved the task if it learns the glossary but damages ordinary prose.
Record each passage once, then send the same recording to every candidate. Don't speak a clearer version for the model that struggled. Include your actual microphone conditions and target language. If several people will use the app, don't assume one developer's voice represents them all.
Keep a human-checked reference transcript for every recording. Before reviewing outputs, decide whether to include false starts in the reference. Don’t quietly remove hesitations from one transcript, then count them as errors in another.
Run an unprompted arm and a vocabulary-assisted arm
Begin by testing without domain hints to establish a baseline. Then repeat the comparison using the vocabulary support available in each chosen implementation. These are separate questions: which recognizer works out of the box, and which works best after a realistic setup? Report both.
Deepgram's keyterm documentation explains terminology prompting for supported models. It also warns that keyterms can appear in usage records. Before sending private names, check the documented redaction option and your data policy. Don't assume an implementation's prompt is a private local dictionary.
For the other candidates, verify that the chosen endpoint currently supports prompting instead of copying a parameter from another API. A Whisper wrapper may offer different controls from the upstream implementation. Save exactly what each request supplied.
Use the glossary from the development set, not the final reference transcripts. Giving the system the expected answer as a hint would make the held-out comparison answer-assisted. Include some unfamiliar terms in the final set to check whether the gains extend beyond the supplied names.
Score speech recognition separately from presentation
Word error rate compares a hypothesis with a reference, counting substitutions, deletions, and insertions. Report your normalization policy alongside the score. Differences in case, punctuation, or how numbers are written can make the texts seem to disagree even when readers understand them the same way.
AssemblyAI's benchmark methodology says it normalizes text before calculating WER. Its published results are evidence for the datasets and settings listed, not a measured ranking for your unpublished dictation sample.
Deepgram's Smart Format documentation says formatting varies by model and language. It can change punctuation and how numbers appear. Record whether Smart Format was enabled. Comparing formatted output with an unformatted reference can lead to an avoidable scoring mismatch.
Keep the original returned text alongside the normalized scoring copy. Review exact names and meaning-changing errors separately. A lower average WER is useful, but it shouldn’t hide a recurring error in the term your users dictate most. The zero-edit benchmark explainer covers finished-message scoring, a different kind of evaluation.
Choose an engine, or decide you need a finished app
Choose self-hosted Whisper when you need local recognition and can support running and maintaining it. Consider model downloads, hardware, and warm-up behavior. The open-source license doesn't remove the hosting or engineering work.
If you want a managed recognition service, shortlist a hosted API. Compare the Deepgram and AssemblyAI configurations that fit your language, deployment, and data requirements. This article doesn't name a price winner or rank services by latency for every use case. Use current quotes and time your own recordings.
If you only need to dictate text into work apps, you may not need to build speech infrastructure. DictaFlow lets you hold to talk on your desktop and insert text at the cursor. Its setup guide explains how it works, and the comparison page compares finished tools rather than raw APIs.
Consumer Pro costs $9/month or $84/year when bought directly for desktop or web, which works out to $7/month billed annually. The free plan includes 2,000 cloud words each month. Product facts distinguish supported Local Offline transcription from online cleanup and list native Windows, Mac, iOS and Android apps. None of this shows that DictaFlow lets you select these three engines.
An accurate transcript doesn't prove the app will insert it correctly. Test insertion separately in the app where you'll use it. For remote sessions, start with the Citrix/RDP insertion guide. Keep the engine comparison and app acceptance test as separate records. That way, you can see what needs to change.
If you're upgrading an API instead of buying a new engine, see the Soniox dictation migration guide. It explains why changing the recognizer and changing the application formatter can be separate release decisions.
Frequently asked questions
Is Whisper more accurate than Deepgram or AssemblyAI?
There’s no universal winner established here. Compare specific models using the same recordings, language, and reference policy. Keep baseline results separate from vocabulary-assisted results.
Can I compare local Whisper with a streaming API?
You can compare the full experiences, as long as you label them as different systems. For recognition-only evaluation, use the same recorded files. Measure streaming finalization and local hardware behavior separately.
Should I give every model my test transcript as a prompt?
No. Use a development glossary, then evaluate on held-out recordings. Supplying the reference transcript gives the test access to expected answers.
Does this comparison report a DictaFlow accuracy score?
No. This is vendor-authored evaluation guidance, not a benchmark we ran. DictaFlow is a finished dictation app, and this article does not claim selectable access to these three engines.