Voice Cloning

Voice cloning lets you create a reusable voice profile from a short reference audio sample. Once a profile is enrolled with a provider, any TTS request can use it to generate speech that sounds like the reference voice instead of a default model voice.

How it works

The cloning workflow has three steps:

1. Create a profile with a name, reference audio file, and reference text (the transcript of what is spoken in the audio).
2. Enrollment happens automatically when the profile is created. Everstack sends the reference audio to the provider, which analyzes the voice characteristics and returns a provider_voice_id.
3. Use the profile by setting voice_clone_profile_id in any TTS request along with the clone-capable model (qwen3-tts-vc-2026-01-22).

After enrollment, the provider retains the voice embedding. You do not need to send the reference audio again on each TTS request.

Provider support

Qwen is currently the only provider that supports voice cloning. Profiles are always enrolled with Qwen regardless of which provider you use for standard TTS.

Profile structure

A voice clone profile stores:

Profiles are scoped to your organization. The uniqueness constraint is (org_id, name), so two different organizations can have profiles with the same name without conflict.

Creating a profile

From the dashboard

Navigate to /deployments/voice in the Everstack dashboard. Click "Create Profile" and provide:

• a name for the profile
• an optional description
• the reference audio file (a short recording of the target voice)
• the reference text (an accurate transcript of what is spoken in the recording)

The dashboard handles the upload, enrollment, and status tracking.

From the API

Use the voice clone profile CRUD endpoints to create, list, update, and delete profiles programmatically. See the Gateway API Reference for the full schema.

Reference audio guidelines

The quality of the cloned voice depends on the reference audio. For best results:

• Duration: 5 to 15 seconds of clear speech. Too short gives the model insufficient signal. Too long introduces noise and variation without improving quality.
• Content: Choose a sentence or two that covers a range of phonemes. Avoid single words or very repetitive phrases.
• Quality: Record in a quiet environment with minimal background noise. A decent microphone matters more than expensive studio gear.
• Format: Standard audio formats (WAV, MP3, FLAC) are accepted. Uncompressed formats preserve more detail for enrollment.
• Reference text: Must accurately match what is spoken in the audio. Mismatches between the audio and text degrade enrollment quality.

Using a cloned voice in TTS

Once a profile is enrolled, reference it in TTS requests:

• Set the model to qwen3-tts-vc-2026-01-22
• Set voice_clone_profile_id to the profile's ID

The gateway resolves the profile, injects the provider_voice_id, and routes the request to Qwen. All other TTS parameters (format, speed, stability, and so on) work the same as with standard TTS.

Workflow studio

The VoiceCloneExecutor node wraps this flow for use inside workflows. It accepts a text input and a voice clone profile reference, then produces audio using the cloned voice. Wire it after any node that produces text, such as an LLM call or a template renderer.

Tenant isolation

Voice clone profiles follow the same tenant isolation model as the rest of Everstack. Profiles belong to one organization, are invisible to other organizations, and cannot be accessed across tenant boundaries. This applies to both the dashboard and the API.

Next steps

• Text-to-Speech covers all TTS parameters and output formats.
• Speech-to-Text explains transcription, translation, and timing data.
• Voice Overview has the full feature summary and provider matrix.
• Gateway API Reference documents the profile CRUD and TTS endpoints.