Skip to content

Make speech and sound

Type a text and have it read aloud in one of nine built-in voices or in a voice of your own, or describe a sound and have it made. Everything runs on your own computer, in the Audio tab; its controls are described in Audio tab. Songs and instrumentals are in Make music.

Kind Model Longest take
Speech Qwen3-TTS 1.7B About 10 minutes (10,000 characters)
Sound Stable Audio 3 Medium 380 s

Download sizes are in Models and licences and what they need to run is in Memory. Cloning and designing a voice each download more files the first time they are used. Stable Audio has licence terms that limit some uses; read them in Licences before you use the output. For a first take, Your first audio goes step by step.

Write the text

Set Kind to Speech and type what should be read in Text, as plain text. It takes up to 10,000 characters, about 10 minutes of speech. Longer text is refused, so split it across takes. A long text is read in parts and joined into one take, every part in the same voice.

Describe it can write the Text and pick a Vibe from an idea, and the script helper can mark up or fit a script you already have.

The model works out the language of the text itself, and the language setting of the model's preset can name one instead. It reads ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.

Insert a pause, a laugh or a delivery

Insert in the Speech tools row, or Alt+Enter in the text, puts a Pause (…), a Laugh (haha) or a mark at the caret. A line break stays a line break inside a part and pauses no longer than … does, so write what you want to hear: … for a pause and haha for a laugh. A delivery mark is not offered while the voice follows a recording.

On a phone: Insert opens as a sheet.

Pick a voice

Voice is the first group in Speech's Settings: your saved voices above the nine default voices. Ryan is the default.

The language of a default voice is its native one, not a limit: every voice reads all ten languages and carries its accent into the others. English text reads without an accent in Ryan or Aiden.

A saved voice is a short reference recording and what it says; any take can read its text in that voice. Saved voices live in voices/ in the data folder (~/.ziv by default) and never appear in the Gallery. A voice comes from a description, a recording or a speech take. The first two are made in the Voices window, which Your voices and New voice, under the dropdown, open; the third is made with Use as in the Gallery or the viewer.

On a phone: Voice is a chip that opens a sheet with the dropdown and the buttons below. The Voices window fits the screen, with its three tabs in a row.

Design a voice

The New voice tab makes a voice from a description, with the model's VoiceDesign files, which download the first time you use it.

  1. Describe the voice (up to 500 characters): age, gender, tone, pace and accent, such as A warm British woman in her thirties, calm and clear. Say how it should sound as well: warm, slow, whispered, excited.
  2. Check the preview line. Every voice reads it, and it is saved with the voice as its transcript, so it must read in 3 to 30 seconds.
  3. Make three candidates and play each one.
  4. Choose one, name it and save it. The voice is chosen for the next takes.

The same description makes a different voice every time, so save the candidate you like: it is a recording that the model clones in every later take, and the voice stays the same. Candidates are kept in voice-design/ in the data folder, never in the Gallery.

Delivery cannot change a saved voice (see Delivery), so put the manner into the description: A calm man in his fifties with a low voice, like a nature documentary narrator.

Clone a voice from a recording

The From recording tab follows a recording of someone's voice. It needs a recording of 3 to 30 seconds of speech, the words spoken in it, and the speaker's consent. Use a clean, close-miked recording without background noise: a phone, radio or old recording makes the cloned voice hiss and sound less like the speaker.

Use once makes it the voice of the next takes; naming and saving it keeps it in Your voices instead. The cloning files download the first time a take follows a recording or a saved voice.

Use a take as a voice

A speech take of 30 seconds or less can become a voice. Open Use as in the Gallery, the viewer or the ⋯ menu of the finished take:

  • Voice opens Speech with the take as a one-off recording and its text under What it says.
  • Saved voice saves it into Your voices, after asking for a name and what it says. A take that followed a recording or a saved voice also asks for consent.

Both are unavailable for a take longer than the model's voice limit. The take must be a speech take; music and sound takes offer neither. Audio track, which gives the take to a video, is not limited; see Use a take elsewhere.

Cloning a voice needs the speaker's permission. Consent is asked per saved voice: a recording, or a take that followed one, asks once when you save it as a voice, and the voice remembers it. Used once, it isn't remembered, so it asks again after a page reload or a change of voice. Designed voices and the default voices are nobody's voice, so they don't ask. The consent is stored in the voice's record in voices/ in the data folder, never in a take's file or the Gallery.

Manage your voices

The Voices window has a tab to manage saved voices (Your voices), one to design a voice (New voice, when the open speech model can) and one to make a voice from a recording (From recording).

To rename or delete a saved voice:

  1. Click Your voices and find the voice on the Your voices tab.
  2. Click its Rename (pencil) or Delete (bin) button. A new name must not be taken by another voice.
  3. Delete asks you to confirm. Takes made with the voice keep their sound.

A voice can't be deleted while a job runs.

Set the delivery

Delivery is how a default voice says the text: emotion, manner (natural, whisper or shout), pace, pitch and expressiveness. The Vibe tiles set all five at once as a starting point, and each stays editable. With every setting neutral, the model reads in its own manner. A take saved with its settings brings the five values back on Reuse.

Delivery works with the default voices only. A saved voice or a recording reads as it was made, so a take that follows one has no delivery. To get a different manner from a saved voice, make a new voice and describe that manner when you design it.

On a phone: Delivery is a chip that opens a sheet with the same Vibes and rows.

Change the delivery partway with marks

A mark in the text changes the delivery from that point on, for a default voice. It is a word in square brackets:

Mark Sets
[sad], [angry], [proud] and the other emotions Emotion
[whisper], [shout] Manner
[as set] Everything back to the delivery you set
I thought we were friends. [sad] You never even called. [as set] Anyway, welcome back.

Insert → Delivery from here puts a mark at the start of a sentence for you. A mark holds until the next mark or the end of the text. It sets one thing and the rest stays as set: [whisper] after [sad] is whispered in the delivery you set, not whispered and sad. The marks are not read aloud and do not count towards the 10,000 characters.

A saved voice or a recording reads the text as made, so its marks are left out.

A bracketed word that is not a mark, such as [sadd], is refused before anything is made, and so is a text of marks alone. Other bracketed text, such as [1], is read as written.

Get help with the script

The script helper works on the script already in Speech's Text box. It marks the script up with delivery, or shortens or lengthens it to a reading time. A script written from an idea comes from Describe it, not from the helper. Script helper in the Speech tools row opens a panel; it is offered wherever the prompt enhancer is set up and is unavailable while a take is being made. It always starts from your own Text, whichever tab is shown.

Action What it does
Fit to time Shortens or lengthens the script so it reads in about the time you give, at the Pace in Settings
Add delivery Adds delivery marks ([sad], [whisper], [as set], …), pauses (…) and line breaks where the meaning asks for them, and nothing else

Pick either or both. Fit to takes 3 to 600 seconds and starts at the script's current reading time rounded to the nearest five seconds. With both picked, one request fits the Text and then adds delivery, so a run makes one version with one record and one Stop. When the delivery finds no place for a mark, the fitted text is kept with a note. Fast or Smart beside the run button is the enhancer. Add delivery is left out while the voice follows a recording.

On a phone: the panel opens as a sheet.

Each result is a version in the Written tab beside your own text. ‹ 2 of 3 › steps through up to five, and you can edit the Written text by hand. Generate sends the Written text. It goes out of date once your own text changes or you pick another model, and is still sent until you scratch it. The script takes at most 4,000 characters.

What the helper keeps, and how a fit is counted
  • Your words. Add delivery never changes a word, a quote mark or a bracket of yours: the result is your own text with only the marks, pauses and line breaks the helper added. Marks and pauses you already wrote stay.
  • Only marks the app knows. It adds only the marks the Insert menu offers (the emotions, [whisper], [shout] and [as set]), at the start of a sentence or a line. A pause is ….
  • Names, numbers, dates and quotes in double quotes are kept as written, and the script is edited in its own language.
  • Fit to time counts the characters that are read, not the marks, at the Pace in Settings. A fit that lands far from the time is kept and says so. A voice that follows a recording is read as recorded, so its time is counted at a Normal pace.

The take keeps your own text and what the helper did: the actions, and with a fit the time and Pace. Reuse settings, Retry and Load settings bring both back, with the script in the Written tab.

Make sound effects and ambience

Set Kind to Sound. The model is Stable Audio 3 Medium, and the same download makes music. Prompt describes what you hear, such as rain on a tin roof, distant thunder. Sound has no Describe it box, since the prompt is already the idea.

In Settings, the Sound section comes first:

  • Type (Auto, Ambience, Effect or One-shot) says what kind of sound it is, and sets the length the sound has when you give none.
  • Length (s) is how long the sound lasts, from 1 to 380 seconds. Typing a number locks it, whatever the type; emptying the field returns to Auto. Reuse settings brings back only a length you gave.

On a phone: Type and Length are chips after Model, each with a sheet.

Polish the prompt with the enhancer

Enhance under the Prompt rewrites it into one short sentence for Stable Audio, at most 45 words, with no brackets, markdown, length or tempo. It describes only what is heard: the source, what it hits, the space, how it changes and how it moves. It adds no music or speech unless you ask for it. The take's Type and Length go along: an effect or a one-shot is written as one event, an ambience as a steady bed, and the rewrite describes only what fits in the take's length.

These options each start on Keep; with all of them on Keep, the enhancer still rewrites your text into the model's shape:

Option Values Key
Space Keep, Small room, Large hall, Street, Open country, Forest, Underwater space
Distance Keep, Close, Medium, Far distance
Over time Keep, Steady, Builds, Hit then tail over_time
Movement Keep, Still, Passing, Approaching, Receding movement
Direction Your own words, up to 300 characters, such as keep the thunder far off note

Reset puts every choice back to Keep and clears the direction. The rewrite is kept in the Enhanced tab beside your text, and Reuse settings, Retry and Load settings bring both back. The enhancer, its Fast and Smart choice and its versions work as they do for images; see Prompt enhancer. The panel opens as a sheet on a phone or tablet.

From the command line

ziv-audio makes the same takes; qwen3_tts reads speech and sa3 makes sound.

ziv-audio -m qwen3_tts --prompt "And then it rained." --speaker aiden
ziv-audio -m sa3 --prompt "rain on a tin roof, distant thunder" --seconds 30

Recipes for saved voices, recordings and delivery are in Make audio.