Last verified: August 30, 2026. Speech API prices and voice cloning requirements were last confirmed against MiniMax documentation on July 5, 2026 and rechecked for this page.
MiniMax Audio is MiniMax’s speech platform. It turns text into spoken audio, clones a voice from a short sample, and gives access to a large library of ready-made voices across dozens of languages. It runs both as a web product and as an API, and the two have different pricing models.
This page is the entry point for everything speech-related on this site: what the platform does, what each model costs, how voice cloning works and what it requires, and which model to pick for a given workload.
Independent-site notice: MiniMax-AI.chat is an independent educational website. It is not MiniMax’s official website and is not endorsed by or affiliated with MiniMax. Product names and trademarks belong to their respective owners. API access, interface availability, billing, and provider terms are controlled by MiniMax.
What MiniMax Audio includes
| Capability | What it does | Where to start |
|---|---|---|
| Text to speech | Converts written text into spoken audio in 40 supported languages | Speech 2.8 |
| Voice library | 300+ system voices plus any custom voices you create | MiniMax Audio guide |
| Voice cloning | Reproduces a specific voice from a short recording | Voice cloning guide |
| Voice design | Creates a new synthetic voice from a description | Voice Design API |
| Multilingual output | Generates the same script across languages and accents | Voice localization |
MiniMax text to speech: what it costs
Speech is billed per million characters of input text, not per minute of audio produced. That is an important distinction when you budget: a dense technical script costs more than a conversational one of the same spoken length, because it contains more characters.
| Model | Tier | Price per million characters | Use it for |
|---|---|---|---|
speech-2.8-turbo | Turbo | $60 | High volume, latency-sensitive work |
speech-2.8-hd | HD | $100 | Published audio where quality is the priority |
To put that in perspective: an average English word runs about six characters including the space, so a million characters is roughly 165,000 words, or about eighteen hours of narration at a normal speaking pace. At $60 that is well under a dollar per hour of finished audio on the turbo tier.
Note that speech is covered by a separate Audio Subscription rather than the developer Token Plan. The Token Plan explicitly excludes Voice Design and Rapid Voice Cloning from its coverage, so a Token Plan subscription does not pay for those. Full pricing across every MiniMax product is on the pricing page.
MiniMax voice cloning
Rapid voice cloning creates a reusable voice from a single recording. The published price is $1.50 per voice, charged once when the voice is created rather than per generation, after which that voice is called like any system voice.
The source recording has firm requirements, and most failed clones are a requirements problem rather than a model problem:
| Requirement | Specification |
|---|---|
| Minimum length | 10 seconds |
| Maximum length | 5 minutes |
| Accepted formats | MP3, M4A, WAV |
| Maximum file size | 20 MB |
Within those bounds, clean input matters more than long input. A clean 30-second sample with no background noise, no music and no second speaker produces a better clone than three minutes of noisy audio. The step-by-step process is in our MiniMax voice cloning guide.
Before you clone a voice, settle the permission question. Cloning a real person’s voice without their consent creates legal exposure in most jurisdictions regardless of what the platform’s terms allow, and the technical ease of doing it is not a measure of whether you may.
Which speech model to use
MiniMax has shipped several speech generations, and older model IDs remain callable. Each generation comes in an HD and a turbo variant: HD prioritises output quality, turbo prioritises latency and cost.
| Generation | Model IDs | Status | Notes |
|---|---|---|---|
| 2.8 | speech-2.8-hd, speech-2.8-turbo | Current | Details — use this for new work |
| 2.6 | speech-2.6-hd, speech-2.6-turbo | Previous | Details |
| 2.5 | — | Previous | Details |
| 02 | speech-02-hd, speech-02-turbo | Legacy | Details |
A practical rule: start on speech-2.8-turbo, generate a representative sample of your actual script, and only move to the HD tier if you can hear the difference on your target playback device. For narration heard on phone speakers, the difference is frequently inaudible and the HD tier costs 67 percent more.
Output formats and settings
Non-streaming requests return MP3, WAV or FLAC. Streaming returns MP3 only. Sample rate, bitrate, output format and channel configuration are all adjustable per request, which matters when the audio has to drop into an existing pipeline with fixed specifications.
If you are wiring this into an application, the request format, parameters and error handling are covered in the MiniMax API guide, and multilingual pronunciation behaviour was tested separately in our Speech 2.8 multilingual test.
Audio in the wider MiniMax range
Speech is one of several MiniMax modalities. Two adjacent points are worth knowing: MiniMax-H3 generates native stereo audio together with video, so a video pipeline may not need a separate speech call at all; and music generation is a separate product line with its own access rules, covered on the Music 3.0 page.
- Hailuo AI — video generation.
- All MiniMax models — the complete list.
- MiniMax API — keys, base URLs and first request.
Prices, model IDs and requirements on this page reflect MiniMax documentation as verified on the dates given. For account, billing and commercial decisions, use MiniMax’s official channels.