Multi-Voice Scripts in Dialogue Editor

Dialogue Editor takes a script with speaker names and returns one audio file with a different voice for each speaker. Its help page explains why that is better than stitching separate Text2Speech clips: it uses ElevenLabs' text-to-dialogue endpoint on the Eleven v3 model, so the whole conversation is produced in one pass, with natural timing between lines. Stitched clips sound like people taking turns at a microphone. One pass sounds like people talking to each other.

The script format

One speaker per line, name first, then a colon, then the words. Here is an invented scene in the format the help page describes:

The county commission meets Tuesday. Two residents got there early.
RUTH: Did you read the new parking plan?
OSCAR: [sighs] All thirty pages.
I even read the appendix.
RUTH: And?
OSCAR: [pause] They're putting meters on the library lot.
RUTH: [laughs] The library lot. Where nobody has ever paid for anything.

The parsing rules matter, because they decide who says what:

• Before the first name: the opening line has no name, so it goes to NARRATOR.
• Continuation: "I even read the appendix." has no name, so it continues Oscar's turn.
• Case: Ruth, RUTH and ruth are the same speaker.
• Blank lines: ignored. Use them freely to keep the script readable.

The counter under the script box shows turns, speakers and characters. Watch the speaker count. If it says four when you wrote three, you have a typo in a name, and "RUHT" is about to get its own voice.

Casting

Every speaker found in the script gets a row in the Cast panel. Click the row, then a voice in the voice browser. Search matches name, accent, age, description and category; filter chips narrow by gender, age, accent and category; the play button plays the voice's preview. Favorites, marked with a star, sort to the top and have their own filter. A speaker shown in red has no voice yet, and Generate will not run until every speaker has one.

Cast for contrast. Two middle-aged American men with similar pitch will blur together in a listener's ear, especially on a phone speaker. Vary at least two of age, pitch, accent and pace. The cast is saved in the browser and comes back next time with the script, which makes a recurring segment easy to keep consistent.

Shaping the delivery

Audio tags

Tags go inside the line: [laughs], [sighs], [whispers], [shouts], [pause] and [sarcastic] are the ones the help page names, and chips under the script box insert them at the cursor. ElevenLabs' prompting guide, which it says applies to Eleven v3 as well as v4, groups tags into voice-related ones such as emotions and delivery, sound effects, and experimental ones such as accents or singing, which it says may be less consistent across voices.

A few habits save credits. Match the tag to the voice, as the guide advises: a soft, breathy voice will not shout convincingly however many [shouts] you add. Do not stack five tags on one line. And give the model enough text to work with. A two-word line gives it almost no context, so test tags inside full sentences.

Stability, seed and quality

Long scripts: split and joined

One request carries up to about 4,500 characters. A longer script is split at turn boundaries, never mid-line, into several requests, and the pieces are joined when you download. The help page is candid about the join: it is a plain MP3 concatenation, it plays correctly everywhere, and the individual parts are offered too. For a clean single-stream file, it suggests running the result through AVPrep with a remux.

A practical rule: if a script is longer than about two parts, break it into scenes yourself, generate each scene, and assemble in SoundStudio. You control the joins, and a fix to scene four does not cost you scenes one through three.

History and the log file

This is the part most people skip and later wish they had not. Every script you generate is kept, with its cast and settings, in the History panel, with no limit on entries. Click an entry and the script, cast, stability, speaker boost, seed and quality come back. "Save to history" keeps a script you have not generated yet, marked as saved. A search box finds scripts by any word, speaker name or voice name and lists the newest 300 matches.

Browser history can vanish when someone clears site data. The log folder is the durable copy. In Chrome or Edge, click Log folder and pick a folder on your computer. From then on every entry is also added to two files there:

• Dialogue Editor log.txt, readable by a person: date, cast with voice IDs, settings, then the script.
• Dialogue Editor log.jsonl, one entry per line, which Import log can read back.

The files are only added to. Nothing in the folder is overwritten or removed. Chrome asks once per page visit whether the editor may keep using the folder; click Reconnect, or just generate or save, and it asks. Anything made while disconnected is marked "not in log" and written next time. Export all downloads the whole history as one .jsonl in any browser, and Import log merges a .jsonl into another browser's history, skipping duplicates. Deleting an entry with the X removes it from the browser only. The log keeps it.

Why care? Because the log answers questions that come later. Which voice did we use for the clerk character in March? What exactly did the script say when we published? Was this line in the version the sponsor approved? Lesson 8 returns to records like this for disclosure and copyright.