Delivery & emotion guide
Qwen3-TTS can follow written instructions such as say it angrily, slowly
, but only in its 1.7B
VoiceDesign and CustomVoice models. These voices are fine-tuned from the 0.6B Base
(voice-clone) model, which has no instruction input and no inline tags: [laugh],
(whispers) or <angry> would be read out as words. Here the emotion comes from the
voice you pick, the punctuation the model learned, and the effects.
What the text controls
| Write | Celebrimbor | C.A.S.S.I.E. |
|---|---|---|
. |
Ends a sentence with a pause. The most reliable way to add a beat: Submit. You are but ash. |
A pause between phrases, the way the game's announcements pause (learned from ~900 game-timed sentences). |
, |
A short breath pause (learned from 181 lines). | Almost never in the training text (1 line), so it may be ignored. Use a full stop for a pause. |
? |
Question intonation, but only weakly (12 training lines). | Not in the training text, so the result is unpredictable. |
! |
Turned into . before speaking, so it has no effect. For force, use the Intense / angry voice. |
Not in the training text. It adds no emphasis, and CASSIE stays flat by design. |
... … |
Turned into . (a normal sentence pause). |
Not in the training text. Use . instead. |
| CAPITALS | Lower-cased before speaking (names such as TALION keep a capital), so shouting in capitals does nothing. | Not learned as emphasis, so don't rely on them. |
| Numbers | Not converted. Write them as words (three hundred). |
Read the game's way: SCP-096 → SCP zero nine six, T-15 → T-minus fifteen, 15 → fifteen. |
Celebrimbor's emotion controls
- Voice is the main emotion switch. Normal (calm, learned from 378 lines) or Intense / angry (shouted commands and shaming, 160 lines).
- SFX stack adds the wraith domination sound: from Dry up to the full boom with the shame sounds and the background. G -6 dB is the default.
- Exact game lines: if the text matches an approved game recording (case and punctuation ignored), the real recording plays instead of the AI. The Voice picks which recording when there are two.
- Another take (on a result) generates again with a new seed. Delivery changes from take to take, so try a few before rewording.
C.A.S.S.I.E.'s emotion controls
- Voice: the plain CASSIE voice, or (with models that have them) the learned announcement voices: Decontamination intro/heavy, Warhead and Countdown.
- Filter rebuilds the game's sound for each announcement type. Pace on Auto slows the line the way the game does (Decontamination heavy 1.3×, Warhead 1.27×, Countdown 1.45×). Slower sounds graver.
- Effects sets the filter strength. In-game effects adds the game's hall reverb or warhead distortion.
- Prefix adds the official intro sounds before the line.
- Source → Real game clips builds the line from the game's own recorded words, so no AI is involved.
- Another take (on a result) generates again with a new seed. Delivery changes from take to take.
Browse word bank ()
Queue & results
Live mic
idleLive mode is off on this server (started with --no-live).
An SFX stack needs the whole phrase before it can play, so each phrase starts a little later than the dry voice.
Continuous: speak; each finished phrase is re-spoken in the chosen voice, phrases queue. Record: speak, press Stop, then play the take. Use headphones so the voice does not hear itself.
Windows Mic Soundboard app: download MicSoundboard.exe · after this one download, use its Check for updates button.
Select text, then press a mood to wrap just that part (each paragraph separately). With nothing selected the mood is inserted at the cursor and lasts to the end of that paragraph. Every paragraph starts as plain narration.
Directives & punctuation guide
Qwen3-TTS only follows written emotion instructions in its 1.7B VoiceDesign / CustomVoice models. Celebrimbor is fine-tuned from the 0.6B Base model, which has no instruction input, so a tag sent to it would be read out as words. The server reads these directives itself and turns them into what it can change in the standard voice: pace (pitch kept), loudness, pauses and the wraith effect (the Intense variant is never used for audiobooks). Laughs, sighs and other sounds that are not speech can't be generated; they become a short pause.
How directives apply
[angry]lasts until the next mood or the end of the paragraph.[angry] … [/angry]covers just that part.- Several at once:
[sad, slow]. Pace and pauses:[slow][fast][pace 1.2](above 1 is slower),[pause 2s][long pause]. - Also understood:
<angry>,{angry},(whispering),*sighs*,[speaking angrily]. Other brackets are read as text; unknown[tags]are skipped with a warning. [chapter The Forge]or a heading line adds a chapter marker (in the MP3 and the timestamps) with longer pauses around it.- A mood switch inside a sentence splits it there, so the delivery changes at that point.
Moods
| Directive | What it does | Also understood |
|---|
Punctuation
. ! | End of sentence: 0.45 s pause. ! is spoken like . (it was never in the training text). Use [angry] (louder, faster) or the option below. |
? | Question, 0.5 s pause. The rising tone is weak (12 training lines). |
... … | At a sentence end: a longer 0.75 s pause. Inside a sentence: a comma pause. |
, | A short breath pause, learned from 181 lines. |
— ; : ( ) | Read as a comma pause. |
| “quotes” | Not spoken. “Go!” he said.is read Go, he said. |
| Paragraphs | A 0.9 s pause. Chapter headings: 1.5 s before, 1.2 s after. The Pauses slider scales them all. |
| Numbers | Read as words: 1984 → nineteen eighty four, 3rd → third, £5.50, 12%, 10:30, 3.5, 1,250. |
| Abbreviations | Mr. Dr. St. e.g. etc. are expanded and don't end the sentence. |
| CAPITALS | Lower-cased (names such as TALION keep a capital): shouting in capitals does nothing. |
| Long sentences | Over 240 characters they are split at the comma nearest the middle. |
Mispronounced names: spell them the way they sound, e.g. Kel-eh-brim-bor
.
Wraith voice = the softer-echo wraith chain on every sentence. [wraith] / [dry] switch it for part of the text.
How it will be read
Paste some text to see each sentence, its voice and the pause after it.
Audiobooks
For other apps: POST /api/audiobook with {"text": "...", "wait": true} returns the audio
links and timestamps; see server/AUDIOBOOK_API.md.