Voice samples
The files behind a cloning voice (the recording, its transcript, affect clips and details), in your voice library or in a character's assets.
A voice-cloning TTS service speaks with a voice sample: a short .wav recording. Everything else about the voice is a file beside it, named after it. The Voice Studio manages these files. This page is for when you edit a folder by hand, assemble a voice pack, or ship a voice inside a character.
The files
For a voice called Anne.wav:
| File | What it is |
|---|---|
Anne.wav | The recording. Only .wav is cloned from. |
Anne.wav.ref_text.txt | Its transcript: the exact words spoken in the recording, as plain text. |
Anne.affect-angry.wav | An affect clip: the same voice said a particular way. |
Anne.affect-angry.wav.ref_text.txt | The affect clip's own transcript. |
Anne.wav.voice.json | The voice's details from the Voice Studio: label, gender, languages, tags. |
Anne.wav.original, Anne.wav.ref_text.txt.original | The recording and transcript before the first saved trim, kept for Restore original. |
The transcript's name is the whole file name plus .ref_text.txt, extension included: Anne.wav.ref_text.txt, not Anne.ref_text.txt. These files belong to the voice. When you move or rename one by hand, move or rename the others with it.
Recordings
Use a short, clean, dry recording of one speaker: no music, no reverb, no other voices. The cloned voice copies whatever is in the clip, room and noise included. A few seconds of natural speech is enough. Some engines refuse long samples (Ming stops at 6 seconds by default).
Transcripts
Most cloning engines (F5-TTS, Orpheus, Sesame CSM, OmniVoice, Ming, audio.cpp) read the transcript alongside the recording. A transcript that matches the recording word for word clones best.
- Write it yourself: put the exact words in the
.ref_text.txtfile, or in the Voice Studio's Transcript box. - Let Voxta write it: leave it out. The service in the Clip Transcription row (Configuration page, Audio tab) writes it the first time the voice is used, then keeps it. One pass covers the recording and every affect clip of it that has no transcript yet. When a chat starts, the voices it will use are transcribed before the language model loads, so the transcription model has the GPU to itself.
Clip Transcription: None means nothing writes transcripts. A voice that has no transcript then fails on the engines that need one, with a message saying so. Pick a transcription service, or write the transcript yourself. Adding Whisper Live, OpenAI, Mistral, Google, xAI or OpenRouter fills the row when it is empty. Other STT services can be picked there.
A transcript the service gets wrong is fixed by editing it: in the Voice Studio, or in the text file.
A character's own voice
A character can carry its voice in its own assets, so it travels with the character in a .voxpak package instead of depending on the user's voice library. Put the files in the character's Assets, under VoiceSamples/:
VoiceSamples/Hero.wav
VoiceSamples/Hero.wav.ref_text.txt
VoiceSamples/Hero.affect-angry.wav
VoiceSamples/Hero.affect-angry.wav.ref_text.txtThen pick it as the character's Voice sample on the Voice tab, where it is listed as Hero.wav (Character Assets). The transcript and affect clip files are not listed. They attach to the voice they belong to.
- Design a voice in the Voice Studio, saved into a character, does this: the recording lands in
VoiceSamples/with the line it spoke as its transcript, and becomes the character's voice sample. - A transcript Voxta writes for a character's voice is saved into the character's assets. It is written once, even when the character is a package or its assets are encrypted. Ship the transcript with the character and nobody needs a transcription service for it.
- When a voice is picked from the library and the character has an asset of the same file name under
VoiceSamples/, the character's own recording is used. - If the sample is deleted, the character falls back to the service's default voice rather than going silent.
Voice packs
Add voices in the Voice Studio takes a zip, laid out flat or with a folder per voice. Transcripts and affect clips in the zip are imported with their voice. You see the plan before anything is written. Entries are skipped when:
- they are not a voice or a transcript (a readme, a cover image),
- they sit in a folder named after another voice (
Elexa/Anna.affect-crying.wav), - they are an affect clip or transcript for a voice that is neither in the zip nor in your library,
- the voice is already in your library (delete it first to replace it),
- they are nested deeper than one folder.