Generate Voiceovers That Laugh, Sigh and Breathe with MiniMax Speech 2.8
Last verified: November 3, 2026
MiniMax Speech 2.8 performs interjection tags that you type straight into a script, such as (laughs), (sighs), (inhale) and (clear-throat). The voiceover reacts the way a person would, on the exact word you choose. MiniMax released this text-to-speech family on January 23, 2026 in two versions, Speech 2.8 HD and Speech 2.8 Turbo. It launched on the MiniMax Open Platform and in MiniMax Audio.
The core of this generation is what MiniMax calls Native Sound Tags. MiniMax modeled everyday fillers like “um,” “uh” and “ah” so that hesitations keep the rhythm, pitch and pauses of real conversation. Among MiniMax's speech models, only Speech 2.8 HD and Speech 2.8 Turbo support interjection tags . The earlier Speech 2.6 pair is listed under legacy models. MiniMax also built this release around studio-grade clarity and better cross-lingual delivery, aiming to remove the “accent bleed” that often affects synthetic speech.
On this page, either version turns a typed script into a finished voice track in a single run.
Explore MiniMax AI's Models
Speech 2.8 Limits and Controls at a Glance
What a Speech 2.8 run accepts, performs and returns.
Script length per run
Up to 5,000 characters in the Text box
Inline interjection tags
(laughs), (chuckle), (sighs), (breath), (gasps), (coughs) and more, on both Speech 2.8 HD and Speech 2.8 Turbo
Preset voices
17 voices, from Wise Woman (default) to Imposing Manner
Emotion styles
Auto (default), Happy, Sad, Angry, Fearful, Disgusted, Surprised, Calm, Fluent or Neutral
Pace and loudness
Speed 0.5–2.0×, Volume 0.01–10 (1.0 = default gain)
Audio file settings
Sample Rate 8,000–44,100 Hz, Bitrate 32,000–256,000 bps, Mono or Stereo
Before You Generate: Six Checks for Cleaner Speech 2.8 Takes
Catch the tag, setting and script mistakes that most often flatten a Speech 2.8 read.
Put each tag where the sound happens
Type the interjection right after the word it reacts to, as in “That was the whole plan (laughs)”, rather than stacking tags at the start of a paragraph. Use the exact lowercase spelling inside half-width parentheses so the cue is read as a sound, not as words.
Keep parentheses for tags only
Speech 2.8 reads half-width parentheses as cues. MiniMax also uses parentheses placed directly after a word to override pronunciation with IPA, Pinyin or Jyutping. Rewrite stage notes and asides such as “(see slide 3)” as spoken sentences, or delete them, so the only words in parentheses are the ones you want performed.
Make Emotion and tags agree
Emotion starts on Auto, which lets MiniMax choose the delivery from your text. If you pick a fixed style, match it to your tags. A (laughs) inside a Sad read pulls the performance in two directions, so use Happy or Auto for playful scripts and Calm for breath-led ones.
Cast the Voice ID before tuning sliders
Pick the preset by role first: Deep Voice Man for a trailer, Lively Girl for an upbeat promo, Elegant Man for a luxury read. Then adjust Speed and Volume around that voice instead of using the sliders to push a mismatched voice into character.
Set the file format for its destination
The defaults (32000 Hz, 128000 bps, Mono) suit everyday web playback. Go up to 44100 Hz and 256000 bps when the file is going into a video or podcast edit, and switch Channel to Stereo only when your delivery spec asks for two channels.
Split long narration at natural breaks
A run takes up to 5,000 characters, and every tag counts toward that total. For longer narration, split the script at paragraph or scene breaks and use the same Voice ID, Emotion and Speed for every part so each section is voiced with identical settings.
Speech 2.8 or ElevenLabs v3: Pick the Engine for Your Script
Some rows cover what each generator on Vidofy offers: the versions, the voice list, how much script one run takes and what it returns. The other rows cover what MiniMax and ElevenLabs publish about their models. Together they show which engine fits your language, your casting needs and how you mark up a script.
| Feature/Spec | MiniMax Speech 2.8 | ElevenLabs v3 |
|---|---|---|
| Versions you can pick | Speech 2.8 Turbo and Speech 2.8 HD, both text to speech | ElevenLabs V3, text to speech |
| Script length per run | Speech 2.8 Turbo and HD: up to 5,000 characters in the Text box | ElevenLabs V3: up to 5,000 characters in the Text box |
| Preset voices to cast | Speech 2.8 Turbo and HD: 17 voices in the Voice ID list, default Wise Woman | ElevenLabs V3: 26 voices in the Voice list, default Rachel |
| Output per run | Speech 2.8 Turbo and HD: one audio file | ElevenLabs V3: one audio file |
| Inline performance tags | Speech 2.8 HD and Speech 2.8 Turbo: interjection tags in round parentheses, e.g. (laughs), (sighs), (breath), (gasps) | Eleven v3: audio tags in square brackets, e.g. [excited], [whispers], [sighs] |
| Supported languages | Speech 2.8 HD and Speech 2.8 Turbo: 40 languages | Eleven v3: 70+ languages |
| Release timing | Speech 2.8: released January 23, 2026 | Eleven v3: introduced as an alpha on June 5, 2025, later made generally available |
How the Differences Play Out in Real Scripts
Moving tagged scripts between engines
Both engines read performance cues written into the script, but the markup doesn't carry over. Speech 2.8 reads cues from round parentheses and Eleven v3 from square brackets, so a script tagged for one has to be re-marked before it goes to the other. Check each cue against the target engine's own tag list instead of assuming the same word exists there.
If you keep one master script for both engines, write the reactions as plain notes. Add the engine-specific tags only in the copy you paste into the Text box.
Language reach, casting depth and version choice
The biggest difference is language coverage. If your script is in a language outside MiniMax's list, Eleven v3 is the one to use. For Chinese, Cantonese, Japanese, Korean, Arabic, Hindi and the major European languages, both engines cover you, and the choice comes down to voice and markup.
On Vidofy, ElevenLabs V3 has a larger preset bench, which helps when a project needs many distinct speakers. Speech 2.8 instead offers two versions, Turbo and HD, with the same voices and controls. That lets one script move between the faster version and the Premium-tier version without recasting.
Which Engine Should Voice Your Next Script?
Use this quick guidance to pick the best option for your workflow.
How to choose: Go with Speech 2.8 when you want two versions on one set of controls, Speech 2.8 Turbo and Speech 2.8 HD; ElevenLabs V3 runs as a single version. Go with ElevenLabs v3 when your script is in a language beyond the 40 that Speech 2.8 covers , since Eleven v3 reaches 70+ . It's also the pick when a cast needs more presets than Speech 2.8's 17, because ElevenLabs V3 lists 26. Both take up to 5,000 characters and return one audio file per run. Both accept inline tags written into the script, but in different markup.
From Script to Tagged Voiceover in Five Steps
Five steps take you from picking a version to a finished audio file.
Choose Speech 2.8 Turbo or Speech 2.8 HD
Select a version in the model list. Both run text to speech and show the same set of controls.
Write the script in the Text box
Type or paste up to 5,000 characters and place interjection tags such as (chuckle) or (exhale) where each sound belongs. If you're starting from a rough idea, Enhance Prompt under the box suggests fuller versions, and your text only changes if you pick one.
Cast the voice with Voice ID and Emotion
Choose one of the 17 presets (Wise Woman is selected by default). Then leave Emotion on Auto or set a style such as Calm, Happy or Surprised.
Tune Speed, Volume, Sample Rate, Bitrate and Channel
Move Speed and Volume away from 1.0 only as far as the read needs, then set Sample Rate, Bitrate and Channel for where the file will play.
Click Generate and listen
Each run returns one audio file. If a reaction lands too early or too late, move its tag in the Text box and generate a new take.