Audio & video editing · Review

Descript Review (2026): The Transcription Is Almost Suspiciously Good

By the AI Tool Court testing desk · Last updated: November 2026 · Protocol v1.0

What Descript actually is

Descript is not a voice generator — it's the editing layer. You feed it recorded audio or video, it transcribes every word, and then you edit the recording by editing the text: delete a sentence from the transcript and it's cut from the audio. It also auto-generates captions, subtitles, and — this is its core business — some of the most accurate speech-to-text in the industry.

We tested it against the job its buyers actually pay for: a podcast or YouTube editor who needs accurate transcripts and captions without hand-correcting every number. That job lives or dies on transcription accuracy, so that's where we aimed the hardest test in our protocol.

Test 1: Transcription accuracy, machine-checked against ground truth

Most reviews say Descript's transcription is "very accurate" and leave it there. We can do better, because we own the answer key. We imported a 47-second AI-generated video whose script we had written — 115 words containing a seven-figure sum, two decimals, four percentages, and the brand names Nginx, Açaí and GIF — then exported Descript's transcript and compared it word-for-word.

Result: 1 genuine error in 115 words (0.9%). The single miss: "processed" came back as "process." Everything else was verbatim-correct once you account for formatting — and the formatting choices were actually smart ones:

Two things worth flagging. First, writing numbers as words is an ASR style choice, not a mistake — every value was semantically exact. Second, spelling NGINX in the brand's official all-caps form is the kind of detail you only get from a transcription engine that has seen a lot of engineering podcasts. This is the best measured accuracy of anything we've tested, and it's consistent with why the tool exists at all: the transcript is the product.

For context across our other tests: Murf scored 18/20 and ElevenLabs 16/20 on text-to-speech accuracy (a different direction — those tools read scripts, Descript writes them down), and Synthesia went 8/8 on a shorter script. Fittingly, it was a Descript-class transcript that we used to verify those tools.

Test 2: The workflow — upload to usable transcript

We dragged a 23 MB MP4 into a new project. Transcription started automatically, and the transcript appeared with speaker labels, paragraph breaks and a word-level caption track already aligned to the timeline. No manual syncing, no correction pass needed given the error count above. Total elapsed time from drag-in to usable transcript: about the length of the video itself.

The free tier also ships "Underlord," an AI editing assistant that suggests first edits (Studio Sound enhancement, auto-captions, fade transitions). We did not score Underlord's suggestions — our panel couldn't consistently judge the Studio Sound difference on our clean test clip, so that feature stays unscored rather than guessed at.

Test 3: What the free tier actually gives you (the part most reviews skip)

The pattern is clear and, frankly, fair: the text outputs are free (transcripts, subtitles — the things that hook podcasters and content teams), while polished video is the upsell. Compared with what we measured elsewhere — Synthesia's free tier allowing roughly one 47-second video with no download at all, and Murf's free tier offering no exports whatsoever — this is the most usable free tier of the four tools we've tested. You can genuinely ship a podcast episode's transcript and SRT captions on $0.

Pricing, mapped from the actual checkout screen (November 2026)

Note Descript bills in local currency with a USD toggle — unusual among the tools we've tested and convenient if your card doesn't like USD charges. At the annual rate, Creator works out to roughly $23/month at current exchange rates for 30 media hours — a fraction of the effective cost of generating equivalent content with AI voice tools, because you're editing audio that already exists.

What we couldn't test

Full disclosure per our protocol: we did not test Overdub (voice cloning) or the filler-word removal workflow at scale, we tested one clean 47-second clip rather than noisy field recordings (Studio Sound's hardest job), and our listening panel could not reliably judge the Studio Sound difference on this clip, so that feature is unscored. We score what we ran; we don't extrapolate.

Who should buy Descript

Buy it if: you produce podcasts, interviews, YouTube videos or meeting recordings — anything that starts as recorded audio. The 0.9% transcript error rate we measured means captions and show notes stop being a correction chore. Creator annual is the sensible tier once video watermarking annoys you.

Skip it if: you need to generate a voice from a script — that's ElevenLabs (realism), Murf (numbers and slides) or Synthesia (talking-head video) territory. Descript edits what exists; it doesn't invent performances. Working on a show? Our podcast intro voice guide covers the generation side.

FAQ

How accurate is Descript's transcription?

In our machine-verified test against a ground-truth script, one genuine error in 115 words (0.9%), with every number and brand name semantically correct. Your mileage will vary with audio quality, but on clean speech it is excellent.

Is Descript free to use?

Yes, with limits: 60 media minutes and 100 AI credits per month. Transcript (.docx) and subtitle (.srt) exports are free; audio exports as lossless WAV; video exports carry a Descript watermark until you upgrade to Creator.

Does Descript have a watermark?

On free-tier video exports, yes — a burned-in Descript mark (our test clip carried both Descript's and the source tool's watermarks). Transcript, subtitle and audio exports have no watermark problem.

Can Descript replace a human transcriber?

For clear audio, at 0.9% error in our test, it gets you a 99%-finished draft instantly — you proofread instead of typing. For noisy multi-speaker recordings, we haven't measured; treat that as untested.

Descript vs Otter.ai — which is better?

We haven't tested Otter yet, so we won't guess. Descript's angle is editing-in-the-transcript plus captions and video; pure meeting-notes transcription is a different job. It's on our bench for the same protocol.

The 0.9% figure above is the transcription half of our benchmark work; the generation side (how well ElevenLabs, Murf and Synthesia read numbers) is tabulated on the benchmarks page, along with the normalization trap that makes most WER claims incomparable.

The numbers, one more time

Tested November 2026 on the free tier, desktop web app, one 47-second 23 MB MP4 import with a ground-truth 115-word script. Transcript diff methodology and rating sheet available on request via editor@aitoolcourt.com.