An in-depth guide to MiniMax Speech 2.8, including 10-second voice cloning, expressive TTS, API pricing and real-time use cases.
Developers building voice products want three things: natural speech, fast response, and the ability to reuse a specific voice. MiniMax Speech 2.8 is the company's flagship text-to-speech family, and it covers all three. It adds expressive "sound tags," low-latency options for live agents, and API-based voice cloning from a very short reference clip.
This guide explains what MiniMax Speech 2.8 offers, how its cloning workflow runs, where it stops short on voice conversion, and what to check before you ship. To compare it with other tools, browse our Voice Cloning & Voice Conversion category.
MiniMax Speech 2.8 is a speech synthesis model family from MiniMax, released in January 2026. It is delivered through an API with two model identifiers:
According to hosting providers, Speech 2.8 improved prosody and naturalness over Speech 2.6 and introduced Sound Tags, a text-injection system for vocal expressions such as laughter and breathing. Third-party model pages also report strong rankings on public blind-test leaderboards, though rankings change over time, so check them before quoting.
The models are also available through partner platforms such as Together AI, Replicate, and Cloudflare Workers AI.
Voice cloning creates a reusable voice from a short reference clip. You then send text and receive speech in that voice. MiniMax documents this workflow clearly.
Voice conversion (a speech-to-speech voice changer that re-speaks existing recordings) is not documented for Speech 2.8 in the sources I found. MiniMax's documented strength is cloning plus text-to-speech. If you need to convert recorded performances into another voice, pair MiniMax with a dedicated voice changer.
| Capability | What it does |
|---|---|
| Rapid voice cloning | Creates a reusable voice from a short reference clip, about 10 seconds according to the announcement |
| Insert tags in text to add laughter, breathing, and other vocal expressions |
| Emotion control | Choose emotions such as happy, sad, angry, fearful, disgusted, surprised, calm, or fluent |
| Speed and volume controls | Adjust pacing and loudness per request |
| HD and Turbo models | Choose between studio-grade quality and low-latency delivery |
| Multilingual speech | Dozens of languages, including Mandarin and Japanese |
| Voice Design | Creates a reusable voice from a text description instead of a recording |
| API and partner access | Use MiniMax's API or hosts such as Together AI, Replicate, and Cloudflare |
Details live in the official MiniMax voice clone documentation.
Speech 2.8 suits developers and product teams who want an API-first voice engine with strong expressiveness and flexible pricing. It is less ideal if you want a polished consumer studio, built-in consent verification, or a voice changer for recorded audio.
| Need | Better fit |
|---|---|
| Expressive TTS and API-based cloning | MiniMax Speech 2.8 |
| Creator studio with a large ecosystem | ElevenLabs |
| Real-time voice agents with ultra-low latency | Cartesia Sonic |
| Open-source, self-hosted cloning | Chatterbox |
| Cloning plus a hosted app and open models | Fish Audio |
1. Confirm you have the rights.
Clone only your own voice or one you have explicit written permission to use.
2. Prepare a clean sample.
MiniMax cites roughly 10 seconds of reference audio. One hosting provider lists 5 seconds as a minimum, and longer samples can improve accuracy. Use a quiet room, one speaker, and natural delivery.
3. Get API access.
Create an account on the MiniMax platform and generate an API key.
4. Upload the audio.
Send the file to the files upload endpoint. The response returns a `file_id`.
5. Create the clone.
Call the voice clone endpoint with the `file_id`, a custom `voice_id`, and a model such as `speech-2.8-hd`. You can also supply prompt text to produce a preview.
6. Use the cloned voice.
Pass your `voice_id` to the text-to-speech (T2A) API. Cloning is billed on first synthesis use, so test deliberately.
7. Tune delivery.
Set emotion, speed, and volume. Add Sound Tags where you want laughter or breathing, and write numbers and dates out fully for more natural speech.
8. Choose the right model.
Use `speech-2.8-turbo` for live agents and `speech-2.8-hd` for final production.
A third-party review reports that, as of August 2026, MiniMax's public documentation describes no enforced consent-verification step for cloning. Under the platform terms, users carry responsibility for having the rights to any voice they clone. That makes your own safeguards essential:
This guide is not legal advice. Consult a qualified professional for commercial or regulated use.
MiniMax uses pay-as-you-go API pricing. Reported figures include:
Partner platforms may price differently, so confirm current rates on the official MiniMax platform before you plan a budget.
MiniMax Speech 2.8 is a strong option for developers who need expressive speech, fast response, and simple API-based cloning. Sound Tags and emotion controls make delivery feel more human, the Turbo model targets real-time use, and the low per-voice cloning fee keeps experiments cheap. Availability through several partner platforms also makes it easy to try.
The trade-offs are important. Voice conversion is not a documented feature, consent verification is left to you, language coverage and sample-length claims vary by source, and there is no polished consumer studio. Start with a short, clean sample of your own voice, test both models on real scripts, and scale only when quality, cost, and compliance meet your standards. Ready to try it? Explore MiniMax.
It is MiniMax's flagship text-to-speech model family, with HD and Turbo variants, expressive Sound Tags, emotion control, and API-based voice cloning. Learn more at minimax.io.
MiniMax cites about 10 seconds of reference audio. Some partner platforms list 5 seconds as a minimum, and longer samples can improve accuracy. See the official voice clone guide.
Voice conversion (speech-to-speech) is not documented for Speech 2.8. Its documented features are text-to-speech and voice cloning.
Sound Tags are text markers you insert to add vocal expressions such as laughter or breathing to generated speech.
HD focuses on high-fidelity audio for final production. Turbo is faster and aimed at real-time uses such as voice agents.
Reported pricing is about $100 per million characters for HD, about $60 for Turbo, and about $1.50 per cloned voice. Check the official MiniMax platform for current rates.
A third-party review says public documentation describes no enforced verification step, so you must secure consent yourself before cloning any real person's voice.
Only with their explicit permission. Using a cloned voice to impersonate or mislead people is unethical and may be illegal.
