MiniMax Speech is a text-to-speech family that grew from the #1-ranked Speech-02 (May 2025) into the current speech-2.8 series. Its strengths are expressive, natural delivery and broad multilingual coverage — 40+ languages with accent and emotion control — plus an unusually cheap way to add custom voices.
Its standout feature is voice cloning: roughly 99% vocal similarity from about ten seconds of reference audio, according to MiniMax, billed as a one-time $1.50 per voice when the voice is first used. Voice Design creates a voice from a text description for $3. A library of 300+ prebuilt voices covers genders, ages and styles. The lineup splits into speech-2.8-hd for studio-grade voiceovers and audiobooks and speech-2.8-turbo for lower-latency, lower-cost use in agents and chatbots.
Pricing
MiniMax’s pay-as-you-go list shows $100 per 1M characters for speech-2.8-hd and $60 per 1M for speech-2.8-turbo, the same for synchronous and asynchronous synthesis. Async jobs accept up to 1M characters per request. That makes MiniMax one of the pricier options per character: Gemini 3.8 Flash TTS is $16.50 and Qwen-Audio 3.1 TTS-Plus $19.30 on Artificial Analysis’s figures, and Eleven v4 is $80 for the higher-quality model. The per-voice cloning fee is the exception — it is the part of the bill where MiniMax can be cheaper than subscription-based rivals.
Where it ranks
On the Artificial Analysis Speech Arena (October 2026), Speech 2.8 HD is #19 (Elo 1174) and Speech 2.8 Turbo #24 (1153), against 1330 for the leader, Eleven v4 Turbo. On the same site’s cloned-voice board, both rank #20–21 at Elo 1020. The family was near the top in 2025, but the September 2026 releases from ElevenLabs, Alibaba and Google have moved ahead.
Who it’s for
MiniMax makes the most sense when you need many cloned or designed voices at a low per-voice cost, or very long asynchronous jobs. For the best quality or the lowest per-character price, other models now do better — see the best AI voice generators. The practical caveats for Western teams: its ecosystem, documentation and tooling are less developed than ElevenLabs’. Model your expected volume before committing.
Best for
- Multilingual voiceover and narration
- Large libraries of cloned or designed voices, where the per-voice fee beats a subscription
- Long-form audio generated asynchronously, such as audiobooks
Pros & cons
Strengths
- Expressive, natural delivery across 40+ languages
- Fast voice cloning (≈99% similarity from ~10s of audio, per MiniMax) for a flat $1.50 per voice, plus $3 voice design from a text description
- 300+ prebuilt voices; HD and low-latency Turbo variants; async jobs up to 1M characters per request
Limitations
- Among the most expensive per character — $100 per 1M for HD against $16.50–$80 for the top five models on Artificial Analysis
- Ranks
- Smaller Western ecosystem and docs than ElevenLabs