MiniMax Speech

MiniMax Speech — Text-to-Speech (Text-to-Audio) Model

What is MiniMax Speech?

MiniMax Speech is a text-to-speech API that turns written scripts into natural, expressive spoken audio. It brings the full MiniMax voice family together behind one endpoint: Speech 2.8, Speech 2.6, and Speech 02, each in an HD tier tuned for quality and a Turbo tier tuned for speed. You send text, choose a system voice, and receive ready-to-use audio in formats like mp3, wav, flac, opus, or pcm.

The model is built for vocal authenticity rather than flat, robotic narration. Across its generations, MiniMax Speech has consistently ranked among the top text-to-speech systems on public blind-listening arenas, making it a strong choice when prosody and naturalness matter.

Key Features

• •Six models, one endpoint: Speech 2.8, 2.6, and 02 in HD and Turbo tiers — switch with a single parameter.
• •Performed interjections (2.8): inline tags such as (laughs), (sighs), (breath), and (clear-throat) are acted out, not read aloud.
• •Pause control: insert timed pauses anywhere with the <#x#> marker (seconds).
• •Emotion control: pick happy, sad, angry, calm, whisper, and more, or let the model auto-select from the text.
• •40 languages: with language_boost to sharpen recognition of a target language or dialect.
• •Fine voice shaping: adjust speed (0.5–2), volume (0.1–10), pitch (-12 to +12 semitones), sample rate, and output format.

Best Use Cases

MiniMax Speech fits narration and voiceover, audiobooks, explainer and product videos, e-learning, podcasts, IVR and announcements, and conversational voice agents. HD tiers suit studio-grade narration where clarity and pacing lead; Turbo tiers suit real-time and high-volume synthesis where latency matters. The 2.8 models are the best pick for expressive, conversational scripts thanks to performed sound tags, while 2.6 excels at reading structured text — URLs, phone numbers, dates, and currency — cleanly.

Prompt Tips and Output Quality

Write scripts the way they should be spoken: punctuation drives rhythm. Use <#0.4#>-style markers for deliberate pauses and reserve interjection tags for the 2.8 models. Match language_boost to your script's language for multilingual or code-switched text, and enable text_normalization when a passage is dense with numbers or dates. Keep speed, vol, and pitch near their defaults for the most natural result, then nudge them per voice. For very long scripts, split text into paragraphs for cleaner pacing.

How to Use MiniMax Speech

MiniMax Speech converts text into natural spoken audio through a single synchronous API call. Send your text, pick a model and voice_id, and the response returns the audio file directly — no polling required.

Choosing a model

• •speech-2.8-hd — highest quality with performed sound tags; best for narration and expressive scripts.
• •speech-2.8-turbo — faster and lighter, still supports sound tags; good for high-volume synthesis.
• •speech-2.6-hd / turbo — strong at reading structured text (numbers, dates, URLs) and the only tier with the whisper emotion.
• •speech-02-hd / turbo — earlier generation, available for compatibility.

Writing the script

Punctuation sets the rhythm. Add timed pauses with the <#x#> marker, for example <#0.4#>, placed between spoken segments. On the 2.8 models you can drop in interjection tags such as (laughs), (sighs), (breath), or (clear-throat), and they are performed rather than read. Match language_boost to your language for multilingual or code-switched text, and turn on text_normalization when a line is heavy with digits, dates, or currency.

Recommended Settings

• •Narration / audiobook: model speech-2.8-hd, speed 0.95–1.0, pitch 0, sample_rate 44100, format mp3 or wav.
• •Conversational / expressive: speech-2.8-hd with inline sound tags and emotion left unset so the model auto-selects tone.
• •Real-time or batch: a Turbo tier with sample_rate 24000–32000 to keep responses fast and files light.
• •Structured info (IVR, announcements): speech-2.6-hd with text_normalization enabled for clean numbers and dates.
• •Fine-tuning voices: keep vol at 1 and pitch at 0, then adjust by ±1–2 only when a specific voice needs it.

Tips

Keep speed, vol, and pitch near their defaults for the most natural delivery, and change one at a time. Choose a voice_id whose language matches your script; for Cantonese voices set language_boost to Chinese,Yue. Use mp3 for web delivery and wav or flac when you plan to edit the audio further. Split long scripts into paragraphs so pacing stays even across the whole passage.

Which model should I use?

Use an HD tier for narration quality and a Turbo tier for speed; choose 2.8 for expressive scripts with sound tags.

Do sound tags like (laughs) work on every model?

No. Performed interjection tags are supported only on speech-2.8-hd and speech-2.8-turbo.

How many languages are supported?

The family covers 40 languages, and language_boost improves recognition of a chosen language or dialect.

How do I add pauses?

Insert a <#x#> marker where x is seconds, for example <#0.5#>, between spoken segments.

What audio formats can I get back?

mp3, wav, flac, opus, and pcm, at sample rates from 8000 Hz up to 44100 Hz.

Is the whisper emotion available on 2.8?

No. The whisper emotion requires a 2.6 model; 2.8 supports the other emotions.