MiniMax voice cloning lets you create a custom voice from an uploaded audio sample and then use the resulting voice_id in MiniMax text-to-speech workflows. The safest way to use it is not to start with the API call. Start with permission, audio quality, and a clear production policy. MiniMax’s official workflow is: upload the source audio, optionally upload short prompt/example audio, call the Voice Clone API, and then use the cloned voice in T2A speech synthesis.
Official sources last checked: July 5, 2026.
This guide explains the practical workflow for developers and product teams: what to check before cloning a voice, how to prepare the source recording, which MiniMax parameters matter, how pricing and retention work, and what to verify before production.
Important: This is a technical and operational guide, not legal advice. Voice cloning can affect publicity rights, privacy, employment rights, platform rules, advertising rules, and fraud risk. Get qualified legal review before cloning or publishing any voice you do not fully own or have explicit permission to use.
Quick Answer: How MiniMax Voice Cloning Works
MiniMax voice cloning is a two-stage workflow. First, you create a cloned voice from a source recording. Second, you use the resulting voice_id in MiniMax T2A speech synthesis to generate new speech.
| Step | What happens | Key output |
|---|---|---|
| 1. Verify consent | Confirm you have rights to use the voice | Written approval / rights record |
| 2. Prepare audio | Record or clean a single-speaker sample | Source audio file |
| 3. Upload source audio | Upload with purpose=voice_clone | file_id |
| 4. Optional prompt audio | Upload short prompt audio with transcript | Prompt file_id |
| 5. Clone voice | Call POST /v1/voice_clone | Custom voice_id |
| 6. Preview and QA | Generate preview or TTS output | Test audio |
| 7. Use in T2A | Pass the cloned voice_id in voice_setting | Final generated speech |
| 8. Govern use | Monitor consent scope, quality, disclosure, cost, and deletion | Production-ready process |
MiniMax’s upload requirements for the main cloning audio are currently mp3, m4a, or wav, with a duration from 10 seconds to 5 minutes and a maximum size of 20 MB.
Who This Guide Is For
This MiniMax Voice Cloning Guide is most useful if you are:
- Building an AI voice agent, avatar, narrator, localization workflow, game character, training module, or internal content pipeline.
- Testing MiniMax TTS with a brand-approved or speaker-approved custom voice.
- Evaluating whether MiniMax voice cloning is production-ready for your app.
- Trying to avoid unsafe or non-consensual cloning practices.
MiniMax voice cloning is not appropriate if you cannot prove permission, if the use case could mislead listeners, or if the voice belongs to a public figure, employee, customer, contractor, family member, or private individual without documented rights and explicit authorization. MiniMax’s own audio terms state that if input content includes an individual’s voice, the user must have full legal rights or appropriate authorization, and MiniMax prohibits impersonation, misleading information, and uses that infringe others’ rights.
Consent and Safety Come Before the API
Voice cloning is not just a technical feature. A cloned voice can sound like a real person, so consent and misuse controls are part of the product design.
MiniMax’s terms say users retain responsibility for input content and must not submit voice data without the necessary permissions. The same terms prohibit impersonating people, misrepresenting affiliation, harming reputation, spreading false or misleading information, and infringing the rights of others.
The broader risk is real: the FTC has warned that harmful voice cloning can be used in scams and that technology alone is not enough to address the risk.
Practical Consent Checklist
Before uploading any voice sample, confirm:
| Consent item | Practical requirement |
|---|---|
| Speaker identity | You know whose voice is being cloned. |
| Written authorization | The speaker or rights holder has approved cloning in writing. |
| Scope of use | The approval states where the voice may be used: ads, apps, narration, internal training, games, support agents, etc. |
| Duration | The approval states how long the voice may be used. |
| Territory and language | The approval covers regions and languages if the voice will be localized. |
| Commercial use | The approval covers monetized or client-facing use if applicable. |
| Revocation process | You know what happens if the speaker withdraws consent. |
| Disclosure policy | You know when listeners should be told the voice is synthetic. |
| Access control | Only authorized team members can generate audio with the voice. |
| Deletion plan | You know how to delete or disable the voice_id if needed. |
What Not to Do
Do not use MiniMax voice cloning to:
- Clone a voice from a video, podcast, call, meeting, social media clip, or leaked audio without permission.
- Impersonate a person in a way that could mislead listeners.
- Create fake endorsements, fake customer service calls, fake HR messages, or fake emergency messages.
- Clone a public figure or celebrity unless rights are verified by qualified counsel.
- Use a cloned voice in political, financial, medical, legal, or employment contexts without additional review.
- Publish cloned-voice content without disclosure when a reasonable listener could be misled.
Audio Requirements and Quality Tips
MiniMax’s official source-audio upload requirements are straightforward: mp3, m4a, or wav; 10 seconds to 5 minutes; up to 20 MB.
But meeting the file requirement does not guarantee a good clone. Quality depends on the recording.
Good Source Audio Looks Like This
| Factor | Recommended practice | Why it matters |
|---|---|---|
| Speaker | One speaker only | Multi-speaker audio confuses the target voice. |
| Room | Quiet, low-reverb room | Echo and background noise can be learned by the clone. |
| Microphone | Consistent mic distance and gain | Reduces volume jumps and tonal inconsistency. |
| Delivery | Natural speech, not exaggerated acting | Produces a more usable general-purpose voice. |
| Background | No music, TV, traffic, keyboard noise, or other voices | Background sound can reduce similarity and stability. |
| Script | Clean, representative phrases | Helps capture pronunciation and style. |
| Format | Use a clean wav or high-quality compressed file when possible | Avoids artifacts from repeated compression. |
MiniMax’s public voice-cloning page also warns that reverb, multiple speakers, and background noise affect cloning quality.
Sample Recording Script
Use a neutral script that covers different sounds, sentence lengths, and pacing. For example:
“This is my approved voice sample for a MiniMax voice cloning test. I am speaking clearly in a quiet room, at a natural pace, using a consistent microphone distance. The final voice will be used only for the approved project described in my consent agreement.”
Do not use sensitive personal details, private information, passwords, medical information, financial information, or confidential company content in the sample.
Pre-Upload Audio Checklist
Before uploading:
- The speaker has given documented consent.
- The recording contains one voice.
- The file is between 10 seconds and 5 minutes.
- The file is under 20 MB.
- The format is
mp3,m4a, orwav. - There is no music, noise, echo, or crosstalk.
- The sample does not include private or regulated information.
- The file name does not expose sensitive personal data.
- You have a rollback/deletion plan.
MiniMax Voice Cloning API Workflow
MiniMax’s official Voice Clone guide describes the workflow as: upload the source audio to get a file_id, optionally upload example/prompt audio, then call the Voice Clone API.
Key MiniMax Terms
| Term | Meaning |
|---|---|
file_id | The ID returned after uploading audio through the file upload endpoint. |
voice_id | The custom identifier for the cloned voice. |
voice_clone | Upload purpose for the main source audio. |
prompt_audio | Optional short sample audio used to improve similarity and stability. |
clone_prompt | Voice cloning parameter object that can include prompt audio and transcript. |
text | Optional preview text for generating a cloned-voice preview. |
model | Speech model used for preview or T2A generation, such as speech-2.8-hd or speech-2.8-turbo. |
language_boost | Optional setting to improve recognition for specific languages or dialects. |
text_validation | Optional transcript used for ASR similarity validation. |
need_noise_reduction | Optional flag to enable noise reduction. |
need_volume_normalization | Optional flag to enable volume normalization. |
MiniMax currently lists speech-2.8-hd, speech-2.8-turbo, speech-2.6-hd, speech-2.6-turbo, speech-02-hd, speech-02-turbo, speech-01-hd, and speech-01-turbo as model options in the Voice Clone API preview field.
Step 1: Upload the Source Audio
The source audio upload uses POST /v1/files/upload with purpose=voice_clone. MiniMax requires Bearer authentication and multipart/form-data for this upload.
curl --request POST \
--url https://api.minimax.io/v1/files/upload \
--header "Authorization: Bearer $MINIMAX_API_KEY" \
--header "Content-Type: multipart/form-data" \
--form purpose=voice_clone \
--form file="@authorized-speaker-sample.wav"Expected result: a response containing a file.file_id. Store that value securely; you will pass it into the voice cloning request.
Step 2: Optionally Upload Prompt Audio
MiniMax supports optional prompt audio. The official docs say providing prompt audio can enhance voice similarity and stability, and the prompt audio upload uses purpose=prompt_audio. Prompt audio currently supports mp3, m4a, or wav, must be less than 8 seconds, and can be up to 20 MB.
curl --request POST \
--url https://api.minimax.io/v1/files/upload \
--header "Authorization: Bearer $MINIMAX_API_KEY" \
--header "Content-Type: multipart/form-data" \
--form purpose=prompt_audio \
--form file="@short-approved-prompt.wav"Use prompt audio only when it adds something useful: a target tone, phrase style, pronunciation, or delivery pattern that the main sample does not capture.
Step 3: Clone the Voice
The Voice Clone API uses POST /v1/voice_clone. MiniMax requires a file_id and a custom voice_id; its docs show optional fields such as clone_prompt, preview text, model, text_validation, accuracy, need_noise_reduction, need_volume_normalization, and aigc_watermark.
MiniMax’s voice_id rules currently require 8 to 256 characters, starting with an English letter, using letters, digits, hyphens, and underscores, not ending with a hyphen or underscore, and not duplicating an existing voice_id.
curl --request POST \
--url https://api.minimax.io/v1/voice_clone \
--header "Authorization: Bearer $MINIMAX_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"file_id": 123456789,
"voice_id": "BrandNarrator_001",
"clone_prompt": {
"prompt_audio": 987654321,
"prompt_text": "This is a calm, clear, approved brand narration sample."
},
"text": "This is a short preview of the approved cloned voice.",
"model": "speech-2.8-hd",
"language_boost": "auto",
"text_validation": "This is my approved voice sample for a MiniMax voice cloning test.",
"accuracy": 0.7,
"need_noise_reduction": false,
"need_volume_normalization": false,
"aigc_watermark": false
}'Use need_noise_reduction or need_volume_normalization as corrective tools, not as a substitute for clean source audio. If the source recording is noisy, the better fix is usually to re-record.
Step 4: Generate Speech With the Cloned voice_id
After cloning, use the voice_id in MiniMax T2A. The HTTP T2A endpoint is POST /v1/t2a_v2; the request body includes model, text, voice_setting, audio_setting, and optional fields such as language_boost, pronunciation_dict, subtitles, and output format.
curl --request POST \
--url https://api.minimax.io/v1/t2a_v2 \
--header "Authorization: Bearer $MINIMAX_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"model": "speech-2.8-hd",
"text": "Welcome to the approved product walkthrough. This synthetic voice is used with permission.",
"stream": false,
"language_boost": "auto",
"output_format": "hex",
"voice_setting": {
"voice_id": "BrandNarrator_001",
"speed": 1,
"vol": 1,
"pitch": 0
},
"audio_setting": {
"sample_rate": 32000,
"bitrate": 128000,
"format": "mp3",
"channel": 1
}
}'MiniMax’s T2A HTTP docs say non-streaming output supports mp3, wav, and flac, while streaming supports mp3; returned URL output, when used, is valid for 24 hours.
Python Example: Authorized Voice Clone Workflow
Use this as a starting point for an internal prototype. It assumes you have consent, a local source audio file, and a MiniMax API key stored as an environment variable.
import os
import requests
API_KEY = os.environ["MINIMAX_API_KEY"]
BASE_URL = "https://api.minimax.io/v1"
headers_json = {
"Authorization": f"Bearer {API_KEY}",
"Content-Type": "application/json",
}
headers_auth = {
"Authorization": f"Bearer {API_KEY}",
}
def upload_voice_clone_audio(path: str) -> int:
with open(path, "rb") as audio_file:
response = requests.post(
f"{BASE_URL}/files/upload",
headers=headers_auth,
files={"file": audio_file},
data={"purpose": "voice_clone"},
timeout=60,
)
response.raise_for_status()
data = response.json()
return data["file"]["file_id"]
def clone_voice(file_id: int, voice_id: str) -> dict:
payload = {
"file_id": file_id,
"voice_id": voice_id,
"text": "This is a short quality-control preview of the approved cloned voice.",
"model": "speech-2.8-hd",
"language_boost": "auto",
"need_noise_reduction": False,
"need_volume_normalization": False,
}
response = requests.post(
f"{BASE_URL}/voice_clone",
headers=headers_json,
json=payload,
timeout=60,
)
response.raise_for_status()
return response.json()
def generate_tts(voice_id: str, text: str) -> dict:
payload = {
"model": "speech-2.8-hd",
"text": text,
"stream": False,
"language_boost": "auto",
"output_format": "hex",
"voice_setting": {
"voice_id": voice_id,
"speed": 1,
"vol": 1,
"pitch": 0,
},
"audio_setting": {
"sample_rate": 32000,
"bitrate": 128000,
"format": "mp3",
"channel": 1,
},
}
response = requests.post(
f"{BASE_URL}/t2a_v2",
headers=headers_json,
json=payload,
timeout=60,
)
response.raise_for_status()
return response.json()
if __name__ == "__main__":
source_file_id = upload_voice_clone_audio("authorized-speaker-sample.wav")
voice_id = "BrandNarrator_001"
clone_result = clone_voice(source_file_id, voice_id)
print("Clone response:", clone_result.get("base_resp"))
tts_result = generate_tts(
voice_id,
"This is an approved synthetic voice sample for internal review."
)
print("TTS response:", tts_result.get("base_resp"))For production, add structured logging, retry handling for transient errors, secret management, request IDs, consent-record references, and review gates before publishing any generated audio.
Pricing and Billing Checks
MiniMax’s official pay-as-you-go pricing page currently lists audio pricing as $60/M characters for speech-2.8-turbo, $100/M characters for speech-2.8-hd, $1.5 per voice for Rapid Voice Cloning, and $3 per voice for Voice Design.
MiniMax’s API overview also notes that using the rapid cloning API does not immediately incur the cloning fee; the fee is charged the first time the cloned voice is used in a T2A synthesis API.
For subscriptions, MiniMax’s audio subscription page currently lists tiers from Starter to Business with different monthly prices, audio points, voice slots, and RPM limits.
Cost Planning Formula
Use this simple planning model:
Estimated monthly cost =
TTS generated characters × selected T2A character rate
+ number of newly activated cloned voices × rapid voice cloning fee
+ preview generations and retry characters
+ subscription, provider, or enterprise costs
+ QA, monitoring, storage, and review overheadAlways verify pricing in the current MiniMax dashboard or pricing page before launch. Pricing, quotas, included credits, provider markups, enterprise terms, and regional availability can change.
Retention, Voice Management, and Deletion
MiniMax’s API overview states that voices produced through the rapid voice cloning API are temporary, and that to keep a cloned voice permanently, you must call a T2A speech synthesis API with that voice within 168 hours, or 7 days.
The API overview adds the operational detail: to keep a rapid-cloned voice permanently, call any T2A speech synthesis API with that voice within 168 hours, or 7 days.
MiniMax’s Get Voice API says voice_cloning voices are available only after being successfully used for speech synthesis.
MiniMax also provides a Delete Voice API for generated voices; the docs state deletion applies to voice_id values generated through the Voice Clone API and Voice Design API, and that once deleted, a voice_id cannot be reused.
Practical Voice Management Policy
For each cloned voice, keep:
- Speaker name or rights-holder reference.
- Consent document ID.
- Approved use cases.
- Approved languages and territories.
- Creation date.
voice_id.- First T2A activation date.
- Owner/team responsible.
- Review status.
- Deletion or expiration date.
- Incident response contact.
Model and Language Considerations
MiniMax’s model overview currently lists speech-2.8-hd and speech-2.8-turbo as current audio models with 40 supported languages, seven emotions, and specified language/dialect support. It lists speech-2.6-* and speech-02-* as legacy models with different positioning.
The T2A HTTP docs expose language_boost, which can be set to a specific language or auto; the docs say it helps recognition for specific minority languages and dialects, and note that speech-01 and speech-02 do not currently support Persian, Filipino, or Tamil.
Which Model Should You Try First?
| Use case | Practical starting point |
|---|---|
| Highest-quality narration or branded content | Try speech-2.8-hd first. |
| Lower-latency or cost-sensitive generation | Test speech-2.8-turbo. |
| Existing workflows using older models | Validate legacy behavior before switching. |
| Long-form content | Consider async TTS workflows after checking current limits. |
| Multilingual output | Use language_boost deliberately and test with native reviewers. |
Do not assume a cloned voice will sound equally natural in every language, accent, script style, or speaking context. For production, test target languages with native speakers and keep a pronunciation regression set for names, brands, acronyms, product terms, and domain vocabulary.
Quality Control: How to Judge a MiniMax Voice Clone
A good voice clone is not just “similar.” It should be usable, stable, understandable, and safe for the intended audience.
QA Scorecard
| Criterion | What to test | Pass/fail signal |
|---|---|---|
| Similarity | Does it sound like the approved speaker? | Approved by speaker or rights holder. |
| Intelligibility | Are words clear at normal listening volume? | No repeated mispronunciations. |
| Stability | Does it stay consistent across scripts? | No random tone shifts or artifacts. |
| Emotion | Does the tone match the use case? | Not too flat, theatrical, or misleading. |
| Language | Does target-language output sound acceptable? | Native reviewer approval. |
| Pronunciation | Names, brands, acronyms, numbers | Regression script passes. |
| Disclosure | Is synthetic use clear where needed? | Listener is not misled. |
| Safety | Could this be used out of context? | Controls exist. |
| Cost | Are previews/retries affordable? | Budget model approved. |
Practical Quality Tips
- Re-record instead of trying to “fix” a bad source sample.
- Keep the source sample conversational, not overly dramatic.
- Use preview text that matches your actual use case.
- Test multiple script lengths.
- Test the voice on laptop speakers, phone speakers, headphones, and noisy environments.
- Use
language_boostonly when it improves the actual target output. - Maintain a pronunciation dictionary or script conventions for recurring terms.
- Get approval from the speaker or rights holder before publishing.
Production Readiness Checklist
Before deploying MiniMax voice cloning in a product or content workflow, confirm:
| Area | Production check |
|---|---|
| Consent | Written consent is stored and tied to voice_id. |
| Rights | Commercial, geographic, language, and duration rights are clear. |
| Access | API keys are stored securely and restricted by environment. |
| Voice IDs | Naming convention avoids personal data and collisions. |
| Retention | You know when unused voices expire and how to delete voices. |
| Pricing | Character costs, voice cloning costs, previews, and retries are budgeted. |
| Rate limits | Current RPM limits are checked before launch. |
| QA | Speaker approval and native-language review are complete. |
| Disclosure | Synthetic voice disclosure is defined for user-facing use. |
| Logging | Requests, outputs, and review decisions are traceable. |
| Safety | Misuse scenarios have been mapped and mitigated. |
| Incident response | You can disable keys, delete voices, and remove published audio. |
| Monitoring | Output quality and complaints are reviewed after launch. |
MiniMax’s rate-limit page currently lists Speech T2A routes at 60 RPM, Voice Cloning at 60 RPM, and Voice Design at 20 RPM, but account-level limits and subscriptions may differ.
Troubleshooting MiniMax Voice Cloning
| Problem | Likely cause | Practical fix |
|---|---|---|
| Low similarity | Poor source audio, wrong speaker, too much noise, weak sample diversity | Re-record a clean sample with one speaker. Try prompt audio if appropriate. |
| Robotic output | Script is unnatural, model mismatch, poor punctuation | Rewrite for speech, test speech-2.8-hd, add natural pauses. |
| Accent issues | Cross-language output needs review | Use language_boost and native-language QA. |
| File upload error | Wrong format, too long, too large | Use mp3, m4a, or wav; keep source audio within 10 seconds to 5 minutes and 20 MB. |
Invalid voice_id | ID violates naming rules or duplicates an existing ID | Use 8–256 characters, start with a letter, and avoid ending in - or _. |
| Preview not returned | Preview text and model may not both be provided | MiniMax says demo_audio is returned if both text and model are provided. |
| ASR validation failure | text_validation does not match recognized source audio | Use an accurate transcript or omit validation if not needed. |
| Voice not listed | Voice may not have been successfully used in speech synthesis | MiniMax says quick cloned voices are available through Get Voice only after successful speech synthesis. |
| Unexpected cost | Preview, T2A usage, or activation fee not accounted for | Recalculate using character usage, preview calls, cloning fee, and plan. |
| API authentication failure | Missing or invalid Bearer token | Use environment variables and confirm the active API key in MiniMax account settings. |
MiniMax Voice Cloning vs TTS vs Voice Design
| Feature | What it does | Best for |
|---|---|---|
| Voice cloning | Creates a voice from uploaded authorized audio | Speaker-approved voices, brand voices, narrator continuity |
| TTS / T2A | Converts text into speech using system or custom voices | Generating final audio from scripts |
| Voice Design | Creates a custom voice from a text description | Fictional voices, character concepts, non-real-person voices |
MiniMax’s API overview describes voice cloning as cloning from user-uploaded audio, while Voice Design generates personalized voices from text descriptions and returns a voice_id that can be used in T2A or async T2A.
Use Voice Design instead of voice cloning when you need a fictional or brand-style voice but do not need to reproduce a real person’s voice.
When MiniMax Is a Good Fit
MiniMax voice cloning is worth testing when:
- You have documented permission to use the voice.
- You need a custom voice for speech synthesis at scale.
- You want API access rather than only a web tool.
- You need multilingual TTS testing.
- You can run quality review before publishing.
- You have a clear deletion and governance process.
Consider another workflow if:
- You do not need a real-person voice.
- You cannot obtain or prove consent.
- Your use case involves high-risk deception, identity, finance, healthcare, politics, or legal advice.
- You need contractual guarantees not visible in public docs.
- You need a provider-specific feature, SLA, region, or deployment model not available in your MiniMax plan.
FAQ
What is MiniMax voice cloning?
MiniMax voice cloning is an API workflow that creates a custom cloned voice from uploaded audio and lets you use the resulting voice_id in MiniMax speech synthesis.
Is MiniMax voice cloning free?
Not necessarily. MiniMax’s current pay-as-you-go pricing lists Rapid Voice Cloning at $1.5 per voice, and the API overview says the cloning fee is charged the first time the cloned voice is used in T2A synthesis. Pricing can change, so verify it in the current pricing page or dashboard before production.
How much audio do I need?
MiniMax’s official upload requirements currently require source audio of at least 10 seconds and no more than 5 minutes.
What file formats are supported?
For source voice clone audio, MiniMax currently supports mp3, m4a, and wav.
Can I clone someone else’s voice?
Only if you have full legal rights or appropriate authorization from the rights holder. MiniMax’s terms say you may not submit voice data without necessary permissions.
Is consent required?
Yes. MiniMax’s audio terms explicitly require full legal rights or appropriate authorization when input content includes an individual’s voice.
Can I use a cloned voice commercially?
Do not assume commercial use is allowed just because the API works. Confirm the speaker’s rights agreement, MiniMax’s current terms, your plan, and any laws or platform rules that apply to your use case.
Does MiniMax support multiple languages?
MiniMax’s current model overview lists speech-2.8-hd and speech-2.8-turbo as supporting 40 languages, and the T2A docs expose a language_boost setting for specific languages and dialects.
Can a cloned voice speak another language?
MiniMax TTS models support multilingual workflows, but you should not assume equal quality across languages. Use language_boost where helpful and test output with native speakers.
What is a voice_id?
A voice_id is the identifier used for a voice in MiniMax speech synthesis. In voice cloning, you define a custom voice_id when calling the Voice Clone API, then pass it into T2A voice_setting.
Does a MiniMax cloned voice expire?
MiniMax says a rapid cloned voice is temporary unless used in T2A speech synthesis within 168 hours, or 7 days.
How can I improve MiniMax voice clone quality?
Use a clean, single-speaker recording in a quiet environment; avoid reverb, multiple speakers, and background noise; optionally use prompt audio when it helps similarity and stability; and run human QA before publishing.
Is MiniMax voice cloning available through third-party providers?
Yes, third-party platforms such as fal, Replicate, Novita, and Together have published MiniMax voice-cloning or MiniMax speech routes. These are third-party access routes, not MiniMax’s official platform documentation, so provider behavior, pricing, data handling, limits, commercial terms, safety filters, consent requirements, and API schemas can differ from MiniMax’s own Open Platform.
What should I check before production deployment?
Check consent, rights, source audio quality, voice_id management, pricing, rate limits, model choice, language QA, disclosure, API-key security, logging, deletion, and incident response.
More on speech: see the MiniMax Audio hub for voices, voice cloning, text-to-speech pricing and which speech model to choose.