Skip to content

Make audio

ziv-audio makes music, speech and sound. The model you pick with -m decides what is made: yue2 and sa3 (with --kind music) make songs and instrumentals, qwen3_tts reads speech, and sa3 makes sound. It needs ffmpeg. The guides Make music and Make speech and sound say what each setting does and why.

Takes are saved as FLAC, named music_, speech_ or sound_ plus the date and time, in the current folder or the one you give with -o, with their settings inside. -r makes several takes with one model load, each with its own random seed unless you set --seed. While a run goes, press N to skip the take, R to repeat it with a new seed, P to pause or Q to quit.

ziv-audio takes no prompt file; to make many takes from a list, use a shell loop.

Music

Make a song with lyrics

ziv-audio -m yue2 --prompt "upbeat synth-pop, female vocal" --lyrics-file song.txt --tempo 96 --key F#m
song.txt
[verse]
Streetlights hum along the avenue
[chorus]
Take me home tonight

--prompt is the style, and the lyrics come from the file, or from --lyrics "..." in the command. You get a song; its length follows the lyrics. --tempo is beats per minute (40 to 220) and --key is C, F#, Dm, F#m… Leave either out and the model chooses. See Write the style and lyrics.

Make an instrumental

ziv-audio -m yue2 --prompt "calm piano, slow" --lyrics "[Intro] [Instrumental] [Solo] [Outro]"

Lyrics with only [section] tags, or none, make an instrumental, and the tags set its length. The command says whether it is making a song or an instrumental before it starts.

Make music with Stable Audio

ziv-audio -m sa3 --kind music --prompt "warm lo-fi beat, soft keys" --tempo 80 --key Dm --seconds 90

Stable Audio takes no lyrics. --seconds sets the length, and left out it is 30 seconds. See Make music with Stable Audio.

Polish the style and make four takes

ziv-audio -m yue2 --prompt "sad folk" --lyrics-file song.txt --enhance genre=folk,instruments=guitar+violin,singer=female -r 4

--enhance rewrites the style into a fuller description once, before the model loads, and all four takes use it. Put + between several values. The keys and values are in the audio enhance spec. --enhance-note "..." is your own direction in words. See Polish the style with the enhancer.

Speech

Without a voice option, the model reads in its default voice, ryan.

Speak in a built-in voice

ziv-audio -m qwen3_tts --prompt "I thought we were friends. [sad] You never even called." --speaker aiden --pace slow --expressiveness high

--speaker picks the voice: ryan, aiden, vivian, serena, uncle_fu, dylan, eric, ono_anna or sohee. The delivery options are --emotion, --manner (natural, whisper, shout), --pace, --pitch and --expressiveness. A mark such as [sad] in the text changes the delivery from there on; see Delivery marks.

To start all five delivery options from a named style, use --vibe, then change one:

ziv-audio -m qwen3_tts --prompt "Good evening, and welcome to the show." --vibe announcer --pitch low

The styles are calm-narrator, storyteller, excited-host, announcer, sad-confession and whispered-secret. The delivery options work only with a built-in voice: they are refused with --voice, because a saved voice or a recording reads as it was made.

Speak in a saved voice

ziv-audio -m qwen3_tts --prompt "And then it rained." --voice Morgan

--voice takes the name of a voice you saved in the Web UI, in any case, or its file slug. A saved voice brings its own transcript and consent, so --voice-text and --voice-consent are refused with it. A file by that name in the current folder is used as a recording first.

Speak in the voice of a recording

ziv-audio -m qwen3_tts --prompt "Welcome back." --voice me.wav --voice-text "Hello, this is me." --voice-consent

The recording is 3 to 30 seconds long, and --voice-text is exactly what it says. --voice-consent confirms that the speaker allows their voice to be cloned; without it the command stops. A speech take made in a built-in voice can be the voice of another take without --voice-consent: give its file as --voice and its text as --voice-text.

The first time a take follows a recording or a saved voice, the cloning files (about 4.5 GB) download after the model's own. The command asks nothing before it downloads. A progress line shows the download, Q stops it, and the licence terms it prints first name only the model's own files. See Clone a voice from a recording.

Read several lines from a text file

Put one line of speech on each line of a file:

lines.txt
Welcome back. It's good to see you again.
The doors open at nine.
Please keep your tickets with you.

A loop makes one take per line, in a folder lines:

while IFS= read -r -u 3 line; do
  [ -n "$line" ] || continue
  ziv-audio -m qwen3_tts --prompt "$line" --speaker aiden -o lines || break
done 3< lines.txt
foreach ($line in Get-Content lines.txt) {
  if (-not $line.Trim()) { continue }
  ziv-audio -m qwen3_tts --prompt $line --speaker aiden -o lines
  if ($LASTEXITCODE -ne 0) { break }
}
while IFS= read -r -u 3 line; do
  [ -n "$line" ] || continue
  ziv-audio -m qwen3_tts --prompt "$line" --speaker aiden -o lines || break
done 3< lines.txt

You get one speech_<date>_<time>.flac per line. Each line loads the model again, so it is slower than -r, which makes several takes of one text. The loop reads the file through a separate input (3<), so the command keeps the terminal and its keys. Press Ctrl+C to stop the loop; a take that fails stops it too, because ziv-audio then exits with 2. See Run commands in a loop and in scripts.

Sound

Make a sound effect

ziv-audio -m sa3 --prompt "a heavy wooden door slams shut in a stone hall" --sound-type effect --seconds 4

--sound-type is auto, ambience, effect or one-shot, and sets the length when you give no --seconds: 12 s for auto and ambience, 5 s for effect and 2 s for one-shot. A length you give is used as it is, from 1 second up to 380.

Make a sound of an exact length

ziv-audio -m sa3 --prompt "wind in the pines, a woodpecker tapping" --seconds 30 --enhance space=forest,distance=far

A sound takes space, distance, over_time and movement in --enhance, with the values in the audio enhance spec. Speech refuses --seconds and --enhance: its length follows the text, and it is read as written. See Polish the prompt with the enhancer.

Flags

Every flag, with its default and limits, is in the command reference.

Flag Does In the Web UI
-m The model: yue2, qwen3_tts or sa3 Model
--kind music, speech or sound, for a model that makes several Kind
--prompt The style of a song, the text to read, or the description of a sound Style, Text, Prompt
--lyrics, --lyrics-file Lyrics of a song (yue2 only) Lyrics
--tempo, --key Tempo and key of a song Tempo, Key
--seconds Length in seconds (music from sa3, and sound) Length (s)
--sound-type The type of a sound Type
--speaker A built-in voice Voice
--voice A saved voice, or a recording Voice
--voice-text, --voice-consent What a recording says, and the speaker's permission What it says, the consent tick
--vibe A named delivery style Vibe
--emotion, --manner, --pace, --pitch, --expressiveness Delivery of a built-in voice Emotion, Manner, Pace, Pitch, Expressiveness
--enhance, --enhance-note, --no-enhance Rewrite a song's style or a sound's description first Enhance, Direction
--enhance-model The enhancer model Prompt enhancer model (Config page)
-r How many takes to make Batch size
--seed The seed, used for every take Seed
-o The output folder Output folder (Config page)