Skip to main content
Reagent turns raw voice recordings into clean, labeled takes. It transcribes the audio, uses an AI model to find each take, then splits, trims and renames the items in REAPER.

Before You Start

  • REAPER is open with the Bridge running and the SWS extension installed.
  • The items are selected. For dialogue editing, each must be 0.5 seconds to 2 hours long. Longer items are rejected, and shorter ones are skipped.
  • You have credits. Transcription is charged by audio length. Analysis, translation and reading a PDF script also use credits (Plans and Credits).
Where your audio goes. Transcription doesn’t run on your computer: Reagent uploads the audio to temporary storage on Reagent’s servers and sends it to ElevenLabs. Script files are uploaded when you attach them. The transcript, and your script if you add one, then go through Reagent’s server to Google’s Gemini model, which finds the takes and reads PDF scripts. The audio isn’t sent to Google. See Privacy & Security.

Automatic Dialogue Editor

Select the items and ask:
1

Transcribe

Reagent transcribes the selected items and detects the language. The Transcribe dialogue step doesn’t ask for approval in any permission mode, and the upload starts as soon as it runs.
2

Choose a mode

Reagent asks whether the recording is Voice Takes or Podcast / Audiobook / Narration, unless your message already says “voice takes”, “pick best takes”, “podcast”, “audiobook” or “narration”.
3

Analyze

Reagent finds and rates the takes in an Analyze transcription step:
  • A breath or hesitation mid-line doesn’t split a take. Starting a line over counts as a retake, even with no silence.
  • Remarks like “let me try that again” are left out of the take text. Other booth talk, such as comments on the performance, becomes a red take.
  • When one voice does at least 70% of the talking in an item, other voices, such as a director’s talkback, aren’t kept as takes. In an interview where the host talks most, this can cut the guest’s lines, so check the result.
  • Each take is rated complete (the whole line was delivered) or incomplete (the speaker stopped early, restarted, or said something that isn’t in the script).
4

Apply edits

Reagent splits each item into one item per take. Each take is trimmed, renamed to its spoken text and given 20 ms fades in and out. Incomplete takes are colored red. Takes keep their timeline positions, and gaps aren’t closed.You approve Apply dialogue edits first, unless your permission mode is Full access.
In podcast mode, or when you added a script, applying edits also deletes items with nothing to keep, including transcribed items that weren’t analyzed or whose analysis failed. Use REAPER’s undo to bring them back. Each edited item is its own undo step, plus one for deleted items.
5

Review the red takes

Listen to the red items. They stay until you ask Reagent to delete them, for example “Delete the incomplete takes”.
Reagent chat editing five selected VO items with a script.csv attached. It reports 60 takes, 18 complete and 42 incomplete takes colored red, and asks whether to delete the incomplete takes.

Recorded in an earlier version of Reagent. The chat now shows the steps described above.

Voice Takes Mode

For sessions where each line is read several times, such as VO or ADR. Reagent finds every take, including false starts, and uses the audio level to place each start and end where the speech begins and ends. Without a script, it reconstructs one from the transcript and rates the takes against it. Add your own to give it the real lines. A take that matches no script line is marked incomplete, colored red, and stays in the project.

Podcast Mode

For one continuous delivery, such as a podcast, audiobook, narration or interview.
  • With a script, Reagent keeps what matches the script, section by section, minus filler words like “um” and “uh”. Longer stretches that don’t match, such as tangents, asides and false starts, become red takes. It also follows editing notes in your message, such as “leave out the sentence about the release date”.
  • Without a script, Reagent keeps the best version of each point. It leaves out filler words, stutters, false starts, repeated attempts and chatter like “let me start over”. Longer stretches it leaves out become red takes. Editing notes aren’t used.
With two or more items selected, Reagent asks whether they are a Single continuous recording or Separate items. If a whole interview is on one recording and you use a script, Reagent keeps the answers. The interviewer’s questions become red takes or are left out.

Add a Script

A script, cue sheet or line list tells Reagent which lines to expect.
  • Formats: CSV, TSV, TXT, PDF, DOCX and XLSX (first sheet only). Markdown (.md) files aren’t used as a script.
  • To add one, attach it (+ > Attach file, drag it in, or paste its full path), or paste the text. More than 10 lines becomes a Script pasted block that Reagent uses as the script.
  • One script per chat. Reagent uses the first script you add, and any supported file you attach counts as one. A later pasted script replaces an earlier pasted one, never an attached file. To switch files, start a new chat.

Pipeline Stages

You can run the stages one at a time, for example to check the transcript first. Only Apply edits changes your project.
Custom instructions. The analysis stage can follow your own instructions instead, such as finding every name, marking questions or spotting topic changes. Reagent can then add markers at the results, split items there, or cut those ranges out.
Set these in your message when you apply edits:
Transcripts and analysis results are saved with the chat on your computer, not in the project. A new chat transcribes again and uses credits again.

Dialogue View

After transcription, ask to see the lines. They appear in a collapsed Dialogue panel (Translation when translated) after a Dialogue manager step. Without translation, the panel shows one line per item, so apply the edits first for one line per take. Showing lines costs nothing extra.
The dialogue view titled "Translation — German to English", with a timestamp, the German line and its English translation on each row, and search, copy and download buttons in the header.

Translation

Reagent can translate transcribed dialogue into another language. Translation runs on Reagent’s servers and doesn’t happen automatically. When the dialogue you edited isn’t in English, Reagent offers to translate it to English. For another language, name it in your message.
Original and translated lines show side by side, and search covers both.

Volume Leveling

The Take volume leveler evens out level differences in dialogue, voice-over and vocal items, and runs on your computer. It writes a take volume envelope on each selected item that turns loud parts down and quiet parts up, replacing the points of any existing volume envelope. Select the items first. Leveling needs REAPER open with the Bridge running and the SWS extension installed, but no transcript. You approve Level take volume first, unless your permission mode is Full access.
You can change any leveler setting in your message.
For an integrated loudness target, say “LUFS”. Reagent levels as usual, then sets each take’s volume so the whole item measures at that value.

Audio Analysis

Find silence, measure loudness and classify sounds.

Example Workflows

Prompts for common tasks, including dialogue editing and leveling.