Best is not a property of a transcription model. It is a property of the match between a model and your particular audio, and that match cannot be read off a public leaderboard. Whisper is open weights you can run on your own machines, which makes it a structurally different choice from a hosted API or a human service before any question of quality comes up. Whether it is still the right pick for you is answered by one hour of your worst recording, not by a ranking.
That sounds like a dodge, so here is what to actually do with it.
The four shapes this decision comes in
Open weights you run yourself. Whisper is the widely used one, and there are reimplementations built for different runtime characteristics, faster-whisper and WhisperX among them. The defining property of this shape is that the audio never leaves your infrastructure and you operate everything.
Hosted transcription APIs. Deepgram, AssemblyAI, Speechmatics and the speech services from the large cloud providers sit here. The defining property is that somebody else operates it and your audio goes to them.
Human and hybrid services. Rev is the familiar name. A person is involved, on a turnaround measured in hours rather than seconds.
Transcription bundled inside something else, where your editing tool, meeting platform or asset system produces a transcript as a side effect. You are not choosing a model here, you are accepting the one that came with the product.
Which shape you are in matters more than which model is currently ahead, because shape determines your data path, your operational burden and who you call when it breaks. What each shape costs varies enough that it is worth checking directly rather than trusting an article, and the shape of that cost is itself part of the comparison.
What a leaderboard cannot tell you
Public benchmarks are run on curated audio. Your audio is a panel with three people talking over each other, a product name nobody outside your company has heard, an accent the training data underrepresented, and a music bed the venue would not turn off.
A model that ranks well on read speech can still make the specific errors that break your workflow. That is why the only comparison worth the afternoon is this one: take your genuinely worst hour, run it through your candidates, and count only the errors that cost something. Proper nouns, numbers, product names, the sentence you intend to quote. Total word error is a research metric. Wrong client name in a published caption is your metric.
The corollary is that if all your recordings are one person in a treated room, most of this deliberation is wasted. Several options will be fine and you should pick on operational grounds.
The failure that no model fixes
If a human listening to the raw audio cannot make out the word, no transcription model will recover it reliably. It will produce something confident and plausible instead, which is worse than a blank, because a blank gets checked.
This is why the highest return move is usually upstream of the model. A lavalier instead of a camera mic, a separate track per speaker, thirty seconds of room tone. What happens when music sits under the dialogue is a good illustration of where the ceiling actually is, and which approach survives noisy source material depends much more on what a wrong word costs you than on which model you started with.
When the model choice barely matters
If you are transcribing in order to find things rather than to publish, the calculus changes. A transcript with scattered errors is still a good index, because you are searching for a passage and you will watch the footage to confirm it. Precision on every word is not what the job needs.
Search over footage also does not have to rest on the transcript alone. Vivu indexes what is visible in the material as well as what is said, so a moment with a garbled word can still come back from what was on screen when it happened. That does not make transcription unnecessary, it makes a transcription error less expensive.
If you are publishing captions, quoting on the record, or feeding a compliance archive, invert all of this. A wrong word is a liability, review time is part of the cost, and a service with humans in the loop stops looking old-fashioned.
How to decide
Answer two questions in order. What does a wrong word cost me, in money or embarrassment? And can my audio leave my infrastructure? Those two answers usually eliminate three of the four shapes before you look at any model.
Only then does the model question matter, and by that point it is not a question you research. It is one you run, on your own hour of difficult audio, and re-run when someone tells you the ranking changed.
FAQ
Should I run Whisper myself or use a hosted transcription service?
The deciding factor is usually not quality, it is whether your audio is allowed to leave your infrastructure and whether anyone on your team wants to operate a service. Self-hosting means you own the machine, the queue, the retries and the upgrade path. A hosted API means somebody else owns those and your audio goes to them.
If you have no constraint on where audio goes and nobody who wants an inference service to babysit, hosted is the lower-friction default. If you have a legal or contractual reason the audio must stay put, that decides it and the rest of the comparison happens inside the self-hosted options.
Does a better transcription model fix bad audio?
No, and expecting it to is the most common way these projects disappoint. If a careful human listener cannot make out the word, the model will not recover it, it will produce a confident plausible substitute, and confident substitutions are harder to catch than gaps.
The gains available upstream are larger than the gains available between models. Separate tracks per speaker, a microphone close to the mouth, and any control you have over room noise will change the output more than swapping models will.
How good is good enough?
That depends entirely on what happens after the transcript. For finding a moment in an archive, a transcript with scattered errors works fine, because you confirm by watching. For published captions or an on-the-record quote, the acceptable number of errors on names and numbers is zero, which means a review pass is part of the workflow no matter which model you chose.
Decide the standard before you test, then measure against that standard rather than against a general sense of quality. Otherwise every candidate looks impressive in the demo and disappointing in production.
Do I need timestamps, or just the text?
Ask this early, because it narrows the field more sharply than accuracy does. If the transcript exists to be read, plain text is enough. If it exists so someone can jump to the moment in the footage, you need word-level or segment-level timing, and you need it to survive whatever tool you import into.
Teams often discover this after processing an archive, and then reprocess everything. Deciding it up front costs nothing.
What about languages other than English?
Coverage varies a lot by language and by accent within a language, and the ranking that holds for English frequently does not hold elsewhere. Nothing published will tell you how a model handles your specific speakers.
Run the same test in the language you care about, using real recordings rather than clean sample audio. If a language matters to your work and a candidate has not been tested on it by you, treat that as unknown rather than as fine.