Verbatim Transcription: Full, Clean and Intelligent

Start with a draft, then apply your convention

In Audio to Text, choose your permitted local recording and select Transcribe this file. English recognition follows a separate language check. Replay the result and edit the words before saving. A fluent sentence is not evidence that the recognizer captured every hesitation. There is no full, clean or intelligent recognition switch on this site; those are review decisions you apply yourself.

Save an untouched draft and retain the recording. Keep a full working copy before making a clean derivative, so an omitted phrase can be recovered. Copy, TXT and Markdown preserve edited text; SRT and VTT require complete valid cue timing. For more about reviewing an exchange rather than a standalone sentence, use the interview transcription guide.

Full verbatim transcription

Under the convention used here, full verbatim retains spoken words, audible fillers, repetitions, restarts and self-corrections. Add observable non-speech events when the agreed specification requires them, using a stated notation. Keep grammatical irregularities as spoken. Do not add a laugh or emotional label merely because the sentence looks hesitant.

Use this level when the assignment explicitly requires those details or when the analysis depends on how a phrase was delivered. A study's written protocol or the recipient's required format takes precedence over this teaching convention. If you cannot verify a passage, mark it unresolved under that specification instead of completing it from context. This tool does not certify an official record.

Clean verbatim transcription

Our clean convention removes an isolated filler and an immediate accidental repetition only when their removal does not erase a response, qualification or correction. It retains the actual proposition, uncertainty, negation and speaker's choice of substantive words. It does not rewrite grammar into polished prose or replace a correction with an inferred final answer.

Use a clean copy when the intended reader needs a readable transcript and the agreed brief permits those omissions. Keep a full copy if the detail may matter later. A repeated word can express emphasis rather than a stumble; decide by replaying the recording and applying the brief, not by deleting every duplicate mechanically.

Intelligent verbatim transcription

The name alone does not define a separate universal standard. Some specifications use intelligent verbatim for the same cleanup as clean verbatim; others permit additional editorial changes. Before choosing it, write down whether you may expand informal wording, omit an abandoned start or summarize a question. An explicit instruction is required for each departure from the spoken wording.

For this exercise, intelligent verbatim follows the same conservative rules as clean verbatim, so there is no separate third file or invented additional edit. If a client or study requires an intelligent-verbatim deliverable, use its supplied definition and resolve conflicts before editing. A summary should be labeled as a summary, not presented as a word-for-word quotation. The terminology differences are visible in the research transcription conventions and the full and clean style explanation.

One recording, two inspectable exports

Download the synthetic source recording, the full verbatim SRT and the clean verbatim SRT. The recording was created for this page with one synthetic English voice. It contains no participant data and is not a real conversational recording. The full text was authored from its own spoken script; it is not an unedited speech-recognition result.

The script starts Um, I, I saved the draft. It then says We can send it Friday, no, Monday. The remaining phrases are I think the label is blue. and Please do not remove the word not. The synthetic source deliberately includes an isolated filler, a repeated pronoun, a corrected day, a qualification and a negative instruction. No emotional state is inferred from its generated delivery.

We constructed phrase boundaries from the actual synthesized audio segments, with silence inserted between them, and loaded that prepared SRT into the existing Subtitle Editor. These are authored phrase windows for this particular recording, not automatic word alignment or measurements of a person's pauses. The editor opened the local WAV and its cue-playback control started the recording. We downloaded the full SRT, changed only the first caption field and downloaded the clean SRT. Both public files are those actual downloads, renamed to identify their convention. The source and method explains the construction.

To reproduce the comparison, save the recording and full SRT. Open or paste the SRT into Subtitle Editor, select Load into editor, then choose Open recording for review. Download the full copy before changing anything. In the first Caption field replace Um, I, I saved the draft. with I saved the draft. Leave the Start and End fields alone. Download the clean copy and compare the saved files, rather than relying on what the editor showed before saving.

Every edit and every unchanged boundary

The first cue removes Um, and one occurrence of I,. Its remaining words are unchanged. Its full and clean time range is 00:00:00,000 → 00:00:03,574. Removing words did not remove audio or shorten that cue. A clean text derivative and an edited recording are different artifacts; this exercise changes only text.

The second cue remains We can send it Friday, no, Monday. in both files, at 00:00:04,574 → 00:00:08,538. The correction is substantive. Deleting no would make the two days look like a list; replacing the whole phrase with We can send it Monday. would hide the original correction. Neither edit is permitted by our clean convention.

The third cue remains I think the label is blue., at 00:00:09,538 → 00:00:11,922. We retain I think because it qualifies the assertion. It is not removed as an automatic verbal tic. The fourth remains Please do not remove the word not., at 00:00:12,922 → 00:00:15,856. Both occurrences of not remain. A cleaner-looking sentence must not reverse the instruction.

All four cue starts and ends match between the actual exported files. The times come from the constructed source segments and the files, not an estimate of recognition performance. Matching timecodes prove only that this text edit preserved the cue windows. They do not prove word-level synchronization, reading-speed suitability or correctness for a different recording. Check the files against this exact WAV; do not reuse its boundaries for a trimmed or resampled recording.

A review rule you can apply to your own recording

Make a short change record for each proposed removal: the original phrase, the proposed derivative, the recording position and the reason the omission is allowed. If the edit removes an answer, a denial, a hedge, a changed name or a correction, stop and check the brief. The full version remains your route back to what was spoken; the clean version is a labeled derivative, not a replacement source.

Before sharing, reopen both exports and compare the affected passage. Confirm that the same speaker is being discussed, the correction still belongs to the same statement and the source recording remains available for authorized review. A hesitation by itself does not establish confidence, dishonesty or a diagnosis. Record observable sounds only; keep your interpretation outside the transcript.

For a plain reading copy, use the SRT to Text converter on the chosen SRT. Removing timecodes does not change the transcription convention. If the words are usable but timing is incomplete, save TXT or Markdown from the transcription result; do not invent end times to enable subtitles. If you need a different timing layout, review it separately against the audio and identify that as a timing revision.

When to arrange human checking

Arrange an accountable human reviewer when the recipient requires one, when a strict convention exceeds what you can verify yourself, or when unresolved speech, overlapping voices or specialist terminology affects the required record. Agree how uncertainty, interruptions and speaker labels will be handled. Payment alone does not settle an ambiguous passage. This page offers a free draft and manual review workflow, not an official transcription service or a promise of admissibility.

The recognition input is English speech in browser-decodable local audio or video, with a 100 MB file-size limit. A maximum supported duration in minutes has not been established. The editor imports SRT, VTT or SBV and exports SRT or VTT; it does not transcribe audio or align plain text automatically. Local files, transcript edits and exports stay in the browser. Model downloads and separate public-page analytics use the network; see the privacy policy.