Voice cloning

A cloning model speaks any text in the voice of a reference clip. The clone copies the speaker, the room and the microphone of the clip: noise, music and a second speaker in the clip come back in every line the voice speaks.

Reference and clone

The reference clip is Narrator A, a synthetic voice made with Kokoro's built-in voice af_heart; no person speaks in the clip. The clone is omnivoice speaking a new line in the voice of the clip, from the shared voice vo_tBeDCRb2byGFjB6HGyQq6a.

The clone of the same clip by every cloning model: Voice models.

The clip

• Before cloning, the silence at both ends of the clip is removed. A clip longer than a model's limit is cut at a pause near the model's cut point (Cloning models). With 20 s of speech, every cloning model except VoxCPM 2 uses the whole clip.
• Send the exact words of the clip as transcript: a cloning model reads the words together with the clip. Without a transcript, the API transcribes the clip on first use and stores the text on the voice. When a model cuts the clip, the model reads the words of the kept part from a transcript of the clip.
• Files up to 20 MB in WAV, MP3, M4A, Ogg, WebM or FLAC. The length comes from the file's header; a WebM clip has duration_s null, and only its first 60 s are used.

Consent

A cloned voice is a person's voice. Clone your own voice, or the voice of a speaker who agreed, and write the agreement in consent: who, for what use, and when. The consent note is visible to the voice's own account only. Never clone a film or TV character or a public figure. A designed voice (POST /v1/media/voices/design) needs no speaker: the model makes the voice from a description.

Check a saved voice

1. Save the voice with POST /v1/media/voices. status is ready at once, and preview_job names the job making the preview.
2. Wait for preview_url and listen to the preview: the voice's default_model reads one sentence in the voice.
3. Generate a paragraph of the text the voice is for. A clone that sounds muffled, echoing or noisy repeats the clip: record the clip again.
4. Each model in the voice's models makes its own clone of the clip: generate the same paragraph on a second model and compare the two files.

The example generates one paragraph in a saved voice on two models. Replace vo_your_voice_id with the voice's id.

VOICE=vo_your_voice_id
TEXT="Thanks for calling the harbour office. The next ferry leaves at \
a quarter past six from pier two, and tickets are on sale at the kiosk \
by the gate until ten minutes before departure."

for MODEL in omnivoice qwen3-tts; do
  curl https://api.albastov.com/v1/audio/speech \
    -H "Authorization: Bearer $ALBASTAR_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"model\": \"$MODEL\", \"voice\": \"$VOICE\",
         \"input\": \"$TEXT\"}" \
    --output "check-$MODEL.mp3"
done

Cloning models

Every cloning model takes voice, reference_text, seed and format. A model uses a clip up to its limit whole; a longer clip is cut at a pause near the model's cut point. Languages and a sample of each model: Voice models.

A saved voice's models lists every model that speaks the voice. OmniVoice also takes instruct, which designs a voice only when the request has no voice.

Reference

• Voices: clone, design, list, copy and delete saved voices.
• Voices reference: every voice route, parameter and error.
• Media jobs reference: each model's parameter schema.