Voice Cloning

Clone voices from reference audio and reuse them with compatible TTS models

Overview

NanoGPT supports voice cloning so you can create reusable custom voices from short reference audio clips and then use them in text-to-speech (TTS).

There are two voice-clone providers exposed via NanoGPT:

• MiniMax voice clone: creates a reusable customVoiceId you can pass as voice when using compatible MiniMax Speech TTS models.
• Qwen voice clone (1.7B): generates a speaker embedding file URL that you can pass to Qwen 3 TTS as speaker_voice_embedding_file_url.

Both flows are asynchronous:

1. Submit a clone job, receive a runId (HTTP 202).
2. Poll the status endpoint until status: "completed".

Authentication

All voice clone endpoints support:

• API key auth: x-api-key: <your NanoGPT API key> (or Authorization: Bearer <key>)
• Session auth (web app): browser cookies

Endpoints

Provider	Submit	Status
MiniMax	POST /api/voice-clone/minimax	POST /api/voice-clone/minimax/status
Qwen	POST /api/voice-clone/qwen	POST /api/voice-clone/qwen/status

MiniMax Voice Clone

Submit a Clone Job

POST /api/voice-clone/minimax

Supports:

• multipart/form-data (upload an audio file)
• application/json (provide audioUrl)

JSON request

{
  "audioUrl": "https://example.com/reference-audio.mp3",
  "customVoiceId": "MyVoice001",
  "voiceCloneModel": "speech-02-hd",
  "needNoiseReduction": false,
  "needVolumeNormalization": false,
  "accuracy": 0.7,
  "text": "Hello! This is a preview of my cloned voice."
}

Form fields

Field	Type	Required	Notes
audio	file	Yes (if no audioUrl)	MP3, M4A, WAV
audioUrl	string	Yes (if no audio)	Hosted audio URL
customVoiceId / custom_voice_id	string	Yes	Must match ^[A-Za-z][A-Za-z0-9]{7,}$
voiceCloneModel / model	string	No	Example values: speech-02-hd, speech-02-turbo
needNoiseReduction / need_noise_reduction	boolean	No	Default false
needVolumeNormalization / need_volume_normalization	boolean	No	Default false
accuracy	number	No	0 to 1, default 0.7
text / previewText	string	No	Preview text

Response (202)

{
  "status": "pending",
  "runId": "abc123-def456",
  "model": "MiniMax-Voice-Clone",
  "cost": 1.0,
  "paymentSource": "USD",
  "isApiRequest": true,
  "fileName": "reference.mp3",
  "fileSize": 245000
}

Poll Job Status

POST /api/voice-clone/minimax/status

Request body

{
  "runId": "abc123-def456",
  "cost": 1.0,
  "paymentSource": "USD",
  "isApiRequest": true
}

Response (in progress)

{
  "status": "processing"
}

Response (completed)

{
  "status": "completed",
  "audioUrls": ["https://cdn.example.com/preview-audio.mp3"],
  "metadata": {
    "model": "MiniMax-Voice-Clone"
  }
}

Qwen Voice Clone (1.7B)

Submit a Clone Job

POST /api/voice-clone/qwen

Supports:

• multipart/form-data (upload an audio file)
• application/json (provide audioUrl)

JSON request

{
  "audioUrl": "https://example.com/reference-audio.mp3",
  "referenceText": "Optional transcript of the reference clip."
}

Form fields

Field	Type	Required	Notes
audio	file	Yes (if no audioUrl)	MP3, OGG, WAV, M4A, AAC
audioUrl / audio_url	string	Yes (if no audio)	Hosted audio URL
referenceText / reference_text	string	No	Optional transcript

Response (202)

{
  "status": "pending",
  "runId": "vc_run_789",
  "model": "qwen-voice-clone",
  "cost": 0.25,
  "paymentSource": "USD",
  "isApiRequest": true,
  "fileName": "audio_file",
  "fileSize": 0
}

Poll Job Status

POST /api/voice-clone/qwen/status

Request body

{
  "runId": "vc_run_789",
  "cost": 0.25,
  "paymentSource": "USD",
  "isApiRequest": true
}

Response headers

While the job is still processing, the response may include an X-Poll-After header indicating how many seconds to wait before polling again.

Response (completed)

{
  "status": "completed",
  "speakerEmbeddingUrl": "https://storage.example.com/speaker-embedding.safetensors",
  "metadata": {
    "model": "qwen-voice-clone"
  }
}

Using Cloned Voices with TTS

MiniMax cloned voice (customVoiceId)

Use your customVoiceId as the normal voice on POST /api/tts with a compatible MiniMax Speech TTS model:

{
  "text": "Text you want spoken in the cloned voice.",
  "voice": "MyVoice001",
  "model": "Minimax-Speech-02-HD",
  "speed": 1
}

Qwen cloned voice (speakerEmbeddingUrl)

Use speakerEmbeddingUrl as speaker_voice_embedding_file_url on POST /api/tts with Qwen-3-TTS-1.7B:

{
  "text": "Text you want spoken in the cloned voice.",
  "model": "Qwen-3-TTS-1.7B",
  "speaker_voice_embedding_file_url": "https://storage.example.com/speaker-embedding.safetensors",
  "reference_text": "Optional: transcript of the original reference audio.",
  "language": "Auto"
}

Saving MiniMax Voice IDs (Web App)

If you use the NanoGPT web app, you can save and list your MiniMax customVoiceId values.

These endpoints are session-authenticated only (they do not support API key auth).

List Saved Voice IDs

GET /api/user/voice-ids

Response

{
  "voiceIds": ["MyVoice001", "MyVoice002"]
}

Save a Voice ID

POST /api/user/voice-ids

Request body

{
  "voiceId": "MyVoice001"
}

Response

{
  "success": true,
  "voiceIds": ["MyVoice001", "MyVoice002"]
}

Voice Clone Storage and Retention

Last verified: February 21, 2026.

Retention depends on the provider behind each voice clone model:

• minimax-voice-clone (WaveSpeed + MiniMax): New cloned voice IDs are temporary. If a cloned voice is not used in a real TTS synthesis call within 7 days (168 hours), it is deleted. If it is used at least once in TTS within that window, it is kept long-term. Preview generated during clone creation does not activate or persist the voice.
• qwen-voice-clone (fal.ai): The returned speaker embedding file URL is hosted by fal. fal guarantees hosted generated files for at least 7 days, then they may be removed at any time. Download and store the embedding yourself immediately for long-term reuse.
• inworld-voice-clone (Inworld Voice API, if enabled in your workspace): Inworld does not publish a fixed auto-delete window for cloned voices in public docs. Treat cloned voices as persistent until explicitly deleted from your workspace. Note: Inworld’s Zero Data Retention mode explicitly does not apply to voice-cloning audio samples.

How to Keep and Reuse Voice Clones

MiniMax / WaveSpeed (customVoiceId)

1. Save the returned voice ID (customVoiceId; provider docs may also call this voice_id).
2. Run at least one real TTS synthesis with that voice ID within 7 days.
3. Reuse the same voice ID in later TTS requests.

Qwen (speakerEmbeddingUrl)

1. Save the returned speakerEmbeddingUrl (speaker_embedding_url in some provider docs).
2. Download the embedding file right away.
3. Store it in your own durable storage (S3, R2, etc.).
4. Use your stored URL later as speaker_voice_embedding_file_url.

Example:

curl -L "$SPEAKER_EMBEDDING_URL" -o my-voice.safetensors

Inworld (voice_id, if enabled)

1. Save the returned voice_id.
2. Reuse it directly for Inworld TTS.
3. If deleted from Inworld, it must be re-cloned.

Can I Download the Clone if It Gets Deleted?

• MiniMax / WaveSpeed: no portable voice embedding download is documented; keep the voice ID active by using it in time.
• Qwen: yes, download the speaker embedding file from speakerEmbeddingUrl / speaker_embedding_url.
• Inworld: no documented voice-embedding export endpoint; keep the voice_id and avoid accidental deletion.

Warning: Provider retention policies may change. This page reflects provider docs as of February 21, 2026.

Provider Source Links

• MiniMax voice cloning intro: https://platform.minimax.io/docs/api-reference/voice-cloning-intro
• MiniMax voice clone endpoint: https://platform.minimax.io/docs/api-reference/voice-cloning-clone
• MiniMax FAQ (voice ID validity, activation, preview behavior): https://platform.minimax.io/docs/faq/about-apis
• WaveSpeed MiniMax voice clone persistence notes: https://wavespeed.ai/docs/docs-api/minimax/minimax-voice-clone
• fal FAQ (file retention): https://fal-d8505a2e.mintlify.app/model-apis/faq
• fal Queue API (X-Fal-Object-Lifecycle-Preference): https://fal-d8505a2e.mintlify.app/model-apis/mndpoints/queue
• Inworld clone voice API: https://docs.inworld.ai/api-reference/voiceAPI/voiceservice/clone-voice
• Inworld list voices: https://docs.inworld.ai/api-reference/voiceAPI/voiceservice/list-voices
• Inworld delete voice: https://docs.inworld.ai/api-reference/voiceAPI/voiceservice/delete-voice
• Inworld zero data retention (voice cloning samples excluded): https://docs.inworld.ai/docs/tts/zero-data-retention

Pricing

Clone runs are charged as a flat per-run fee:

• MiniMax voice clone: $1.00 per run
• Qwen voice clone (1.7B): $0.25 per run

The submit response includes cost and paymentSource for the run.

Limitations

• MiniMax and Qwen clone endpoints are asynchronous; clients must poll status until completion.
• MiniMax customVoiceId must match ^[A-Za-z][A-Za-z0-9]{7,}$.