AI models on Craze
Every video, image, voice, music and sound model Craze can direct, with what each one makes and what it costs.
Every video, image, voice, music and sound model Craze can direct, with what each one makes and what it costs.
Starts at$0.106 per 1,000 characters
MiniMax's high-definition TTS with expressive sound tags and voice selection. Write a script, optionally choose a voice ID, and Craze narrates.
Craze runs MiniMax Speech 2.8 HD when you ask for it by name.
This route does not accept audio reference inputs.
Mark up your script with tags like (laughs), (whispers), (angry), or (breathes).
The model prioritizes sounding human.
Switch languages mid-script without regenerating.
Craze creatively jams, researches and plans with you.
Craze generates anything imaginable. Video, audio, music, motion graphics and more.
Craze watches your footage and builds a final cut. With a real editor if you need it.
Every model here runs on every Craze plan, including Free, and Craze can use a different one for each shot.
| MiniMax Speech 2.8 HDMiniMax | ElevenLabs v4ElevenLabs | Gemini 3.8 Flash TTSGoogle DeepMind | Gemini 3.8 Flash Lite TTSGoogle DeepMind | |
|---|---|---|---|---|
| Makes | Text to audio | Text to audio | Text to audio | Text to audio |
| From | $0.106 per 1,000 characters | $0.085 per 1,000 characters | $0.032 per 1,000 characters | $0.022 per 1,000 characters |
| Task | Inputs | Resolution | Aspect ratios | Sound |
|---|---|---|---|---|
| Text to speech | Text prompt | Follows the source | Follows the input | None |
$0/ month
$19/ month
$59/ month
$199/ month
MiniMax Speech 2.8 HD is a high-definition text-to-speech model that combines voice selection with expressive sound tags. It treats narration like a performance, not a robotic read-aloud.
Speech 2.8 HD delivers studio-quality narration by combining naturalness, emotional delivery, and sound tags without separate tracks. The ability to add laughs, sighs, and other non-speech sounds inline makes it stand out for character work and narrative-driven projects.
This route does not accept audio reference inputs. Select one voice ID for a generation; generate separate character lines with different voices if needed.
Mark up your script with tags like (laughs), (whispers), (angry), or (breathes). The model generates those sounds as part of the narration, not as separate stems you stitch together. A character can laugh, then speak (both from one generation).
The model prioritizes sounding human. Inflection, breathing, and timing feel intentional rather than synthetic. That matters for character work, where a flat voice breaks immersion.
Switch languages mid-script without regenerating. A sentence in English, a word in French, back to English. The voice adapts to each language and stays consistent across the switch.
(warmly) or (shocked) and non-speech sounds like (door slams).Pricing starts at $0.106 per 1,000 characters and varies with script length. This route has no separate price axis for voice selection, sound tags, or language. Use the calculator to estimate cost.
Games, animations, and short films often need multiple distinct voices. Select a voice for each character's generation and assemble the lines in your project.
Tutorial videos, audiobooks, and branded content benefit from laughter, gasps, and emotional inflection. MiniMax lets you write (laughs nervously) and have that sound emerge naturally from the narration.
If your script spans languages, MiniMax handles code-switching in one pass. A global brand narrator can deliver copy in English, Spanish, and Mandarin without three separate generations.
Seed Speech from ByteDance prioritizes cost-effective narration with optional voice selection. MiniMax Speech 2.8 HD offers integrated sound tags and 32-language support in a single generation.
For expressive English narration with steerable prompts, Gemini 3.1 Flash TTS gives you directorial control through inline tags. Seed Audio 1.0 accepts audio or image references for speech. For emotion-tagged narration across 16 languages, Qwen Audio 3.0 TTS is another option.
MiniMax Speech 2.8 HD supports emotion tags like (angry), (happy), (sad), (surprised), and (nervous), plus non-speech sounds such as (laughs), (sighs), (whispers), (coughs), and (breaths). You can stack tags for nuance: write "I can't believe it" marked as (shocked, breathless) for a more specific emotional read. The model maintains consistency across tags within the same generation, so a character who sounds angry early will sound like the same person when they're surprised later, not a completely different voice.
On Craze, each sound tag generation takes 5-20 seconds depending on script length. If your script is very long with many sound tags, break it into sections for faster iteration. Test a short section first to dial in the tone before committing to a full-length generation. MiniMax includes preset emotional profiles you can apply to a whole voice to shift the baseline mood before adding inline tags.
This route accepts no voice sample. To use an existing voice, provide its voice ID; a separate voice-profile workflow is needed to create one from a recording. Generate each voice separately for a multi-character scene.
Regenerating the same script with the same selected voice may produce a different take; review each result for delivery and consistency.
When you switch languages mid-script with MiniMax, the sound tags adapt to the language context. A (laughs) tag in English sounds the same as in Spanish or Mandarin, but the voice's language-specific accent and timing shift automatically. This is valuable for global brand work where you want one voice across markets but need localized narration.
MiniMax Speech 2.8 HD accepts up to 4,000 characters of text per request on Craze. For longer scripts, split the text and combine the results in your timeline. The model uses punctuation to infer pacing and emphasis, so add commas and periods to control the reading pace.
Voice cloning is not an input mode for this Craze TTS route. It accepts a voice ID but no reference audio.
Describe what you want the way you would brief an editor, and Craze writes the MiniMax Speech 2.8 HD prompts for you. Every Craze plan, including Free, runs MiniMax Speech 2.8 HD and every other media model, and Craze can use a different one for each job. Each voiceover lands on the timeline under your cut, and exports with the video or on its own as MP3. Generations come out of your plan's monthly credits, from $0.106 per 1,000 characters, with top-ups when you need more.
Character voice work and dialogue: Games, animations, and short films often need multiple distinct voices. Expressive narration with ambience: Tutorial videos, audiobooks, and branded content benefit from laughter, gasps, and emotional inflection. Multilingual scripts and global brands: If your script spans languages, MiniMax handles code-switching in one pass.
MiniMax Speech 2.8 HD generates narration with expressive sound tags like (laughs) and (sighs). The Craze route takes text and an optional voice ID, not a voice sample.
MiniMax released Speech 2.8 HD on January 23, 2026, with native support for sound tags and zero-shot voice cloning from short reference audio.
You pay $0.106 per 1,000 characters. The cost calculator on this page shows the total before generation.
every Craze plan, including Free.
Not in this text-to-speech route. It accepts text and an optional voice ID, not an audio sample. Use a separate voice-profile workflow to create a voice first.
Sound tags like (laughs), (sighs), (whispers), and (angry) let you add non-speech sounds and emotional direction inline in your script. The model generates them as part of the same narration pass.
Yes. It supports 32 languages and can switch between languages within the same script. Voices adapt to each language automatically.
