Read and verify a transcript result
What the transcript screen shows, why RoughCut flags time ranges instead of dropping them, and how to check the flagged parts.
The result screen opens with a Your transcript is ready banner and two exports, Export SRT… and Export Text…. A statistics row shows the word count, the duration, the detected language in uppercase such as EN or AR, and the number of cues. The transcript itself is a cue list — the same blocks, in the same order and with the same timing, that the SRT export contains — with a play button on every cue and a couldn't hear cleanly pill on any cue the model could not hear cleanly. Where RoughCut could not transcribe a region cleanly it keeps that audio and flags the time range under a header that counts the flagged sections and tells you to listen before trusting them; a flagged section is not necessarily wrong. Known whisper hallucination phrases — subtitle credits, canned sign-offs over music or room tone — are removed from transcribe-only results and listed with their timecodes, so nothing disappears silently. Each cue's text is selectable and follows the direction of its own words, so Arabic renders right to left. RoughCut does not include a full transcript editor in the current version.
On this page
What the screen shows
A finished Transcribe Only job opens on its result. The banner reads Your transcript is ready, followed by two export buttons — the prominent Export SRT… and the secondary Export Text… — a statistics row, any flagged ranges, a note listing any known hallucination phrases that were removed, and then the transcript as a list of cues.
The screen's own line under the banner states the design principle:
Export it as SRT subtitles or plain text. Anything the model couldn't hear cleanly is flagged
below, and any known whisper hallucinations it removed are listed — nothing disappears silently.The statistics row
RoughCut shows four figures. There is no confidence score, no speaker count, and no accuracy percentage, because the app does not measure those.
| Statistic | What it is | Note |
|---|---|---|
| Word count | The number of words in the transcript | Counts words, not characters or lines |
| Duration | The length of the transcribed media, in minutes and seconds to one decimal | Formatted like 12:04.6 |
| Detected language | The language the transcriber identified, uppercased | For example EN or AR |
| Cue count | How many subtitle blocks the transcript was split into, shown as <n> cues | The same count the SRT export will have |
The detected language is what the transcriber concluded, not what you asked for. If you left Language on Auto-detect and this reads the wrong language, run the file again with English or Arabic forced. See Switch between English and Arabic for the difference between the job language and the interface language.
Flagged ranges, not silent deletions
This is the part worth understanding, because it is the opposite of what most transcription tools do.
When RoughCut is unsure about a stretch of audio, it keeps whatever it produced for that stretch and flags the time range. It does not quietly delete the passage, and it does not fill the gap with a confident-looking guess and move on. A warning block appears above the transcript with a header of the form:
3 sections flagged — listen before trustingBelow the header, each flagged range is listed on its own line with its start and end timecode and a short reason. A single flagged range uses the singular wording, 1 section flagged — listen before trusting.
The whole point of the block is to send you to specific timecodes. The cues inside a flagged range carry a couldn't hear cleanly pill and a faint tint, so you can spot them while reading, and each has its own play button. Listen, then correct the exported transcript by hand where it matters.
The two flag kinds
Two different reasons can put a range in that list. They come from different checks, but they ask exactly the same thing of you.
| Reason shown | What produced it | What you do |
|---|---|---|
couldn't transcribe cleanly | The stronger signal. RoughCut watches for hallucination-style output — a phrase repeating over and over — and for words that sit almost entirely inside a detected silence. Either pattern marks the range as unreliable. | Listen to the range. Expect the text there to be repeated, invented, or absent. |
transcription may have degraded | The weaker signal, raised by the transcriber's own health check when output quality appears to have dropped over a stretch. | Listen to the range. It is often fine; treat it as a prompt to check rather than a verdict. |
Both lines are printed with the range first and the reason after, in monospaced digits so the timecodes line up.
A flag is not a verdict
Flagged sections are not necessarily wrong. The flags mark places where RoughCut's confidence in its own output dropped, which is a different claim from "this text is incorrect". Plenty of flagged ranges transcribe perfectly well; some unflagged ranges will still contain errors, because no transcriber knows everything it got wrong.
Treat the list as a triage queue in priority order, not as an error report. If you are producing captions or quoting the transcript, listen to every flagged range before publishing. See Export SRT captions.
The same underlying flags appear in a rough cut too, where they are attached to the affected segments in the review list rather than listed as bare timecodes. See Review cuts on the timeline.
Known whisper hallucinations are removed, and listed
Whisper-family models have a small repertoire of phrases they produce over music, silence, or room tone that were never spoken: subtitle-credit lines, and canned sign-offs of the thanks-for-watching and subscribe-to-the-channel kind. On a Transcribe Only job, RoughCut removes them.
The rule is deliberately narrow. A credit line is dropped wherever it appears. A sign-off phrase is dropped only when the audio underneath it is essentially quiet — say it into the microphone and it stays. And the removal is never silent: a note above the transcript counts what went, <n> known whisper hallucinations removed, lists each phrase with its start and end time, and explains:
These phrases are whisper artefacts over music or room tone, not speech. They are left out of
the SRT and text exports.Rough cuts are not touched by this step; their transcripts go through the editor pass and the review list instead. If a recording consists of nothing but such phrases, the job fails rather than producing an empty transcript, with transcription produced only known hallucination phrases — no speech to keep.
Reading and exporting the transcript
The transcript is shown as cues: the same blocks, in the same order and with the same timing, that Export SRT… will write. Each cue shows its start and end time, a play button that auditions exactly that range of the source audio, its text, and — where the range was flagged — a couldn't hear cleanly pill. What you read on screen is block for block what the SRT contains, so a cue that reads oddly here will read oddly in your captions.
Each cue's text is selectable, so you can copy a passage straight out of the window without exporting. Its direction follows the words themselves, not the app's interface language: an Arabic cue renders right to left even while RoughCut's interface is in English, and an English cue renders left to right in the Arabic interface.
Two exports are offered:
Export SRT…writes subtitle blocks withHH:MM:SS,mmmtimecodes, saved as<stem>.srt.Export Text…writes the plain transcript, saved as<stem>.txt.
Both are described in Supported inputs and generated files.
What is not here
RoughCut does not provide a full transcript editor in the current version. You cannot retype a word, merge a caption block, or fix a flagged range inside the app and re-export a corrected file. The workflow is: export, then correct in whatever editor you already use — a text editor for the TXT, or your captioning tool for the SRT.
There is also no speaker labelling, no per-word confidence display, and no translation. The transcript is in the language that was spoken.
Common mistakes
- Reading the flag count as an error count. It is a count of ranges worth checking, not a count of mistakes.
- Exporting captions without listening to the flagged ranges. These are exactly the passages most likely to embarrass you on screen.
- Assuming an empty flag block means a perfect transcript. No flags means no check fired, not that every word is right.
- Trying to fix the text in the app. Export first, then edit.
- Looking for a removed hallucination phrase in the exports. It is left out of both, by design; the list on screen is the record.
- Blaming detection for a wrong language. Force
EnglishorArabicon the job and run it again.
Troubleshooting
| Symptom | Likely cause | What to do |
|---|---|---|
Loading transcript… never finishes | The job folder is being read | Wait; if it persists, reopen the job from the sidebar. |
Could not load the transcript: … | The transcript file could not be read | The job folder may have been cleaned up. Finished jobs are removed after the retention window — see Manage models, data, jobs, and storage. |
SRT file is missing from the job folder. | The SRT artifact is not present | Re-run the file. If the job folder was cleaned automatically, the artifacts are gone with it. |
| Most of the file is flagged | Very noisy or very quiet audio | Transcribe a separate microphone track if you have one, and check the Recording environment choice. See Match the recording environment. |
| The transcript repeats one phrase for a long stretch | Hallucination loop in difficult audio | The range should be flagged as couldn't transcribe cleanly. Cut that passage from the exported text and listen to what was actually said. |
| The detected language is wrong | Auto-detection guessed badly, often on short or mixed-language clips | Re-run with Language set explicitly. |
transcription produced only known hallucination phrases — no speech to keep | The whole file was music, silence, or room tone that the model filled with credit lines and sign-offs | There is no speech to transcribe. Check that you chose the right file, or the right audio track. |
Still stuck? See Troubleshooting by symptom.