Turn Raw Scripts into Natural Speech with MiniMax Speech 2.6
Last verified: November 3, 2026
MiniMax Speech 2.6 is MiniMax's text-to-speech family for voice-agent work. It was released on October 29, 2025 in two versions, Speech 2.6 HD and Speech 2.6 Turbo. What sets it apart is how it handles the hard parts of real scripts. It converts URLs, email addresses, phone numbers, dates and money amounts straight into speech, in multiple languages, so each line is read correctly from the start . You can paste a helpdesk message, a sale announcement or a museum guide as it is and get a natural, human-like read without rewriting anything phonetically first.
For this release, MiniMax rebuilt its audio generation pipeline to get end-to-end latency under 250 milliseconds, so speech generation is no longer the slow part of a live conversation . Prosody also moves closer to natural human delivery. The release adds Fluent LoRA, which builds on Speech 2.5's voice cloning, which already kept a speaker's accent and speech habits. Fluent LoRA keeps a voice's timbre even when the source recording is accented or disfluent, and it still produces fluent speech that matches the target text .
MiniMax Speech is also used in production voice-agent tools. MiniMax names LiveKit, the open-source framework Pipecat and the voice platform Vapi among the builders that chose it as their engine . On this page, Speech 2.6 Turbo and Speech 2.6 HD share one script box and the same voice, emotion and audio controls.
Explore MiniMax AI's Models
Speech 2.6 Turbo & HD at a Glance
What you can enter, set and get back from Speech 2.6 Turbo and HD on each run.
Versions & output
Speech 2.6 Turbo or Speech 2.6 HD; one audio file per run
Script length
Up to 5,000 characters per run
Voice presets
17 voices, from Wise Woman (default) to Imposing Manner
Emotion styles
Auto, or Happy, Sad, Angry, Fearful, Disgusted, Surprised, Calm, Fluent, Neutral
Speed & volume
Speed 0.5–2.0× (default 1); Volume 0.01–10 (default 1)
Audio file quality
8,000–44,100 Hz sample rate; 32–256 kbps MP3 bitrate; mono or stereo
Check These Before You Generate
Six quick checks that prevent stilted reads, mismatched parts and audio files that don't fit where they'll play.
Paste links, numbers and contact details as written
Speech 2.6 turns URLs, email addresses, phone numbers, dates and money amounts into speech by itself. Spelling them out by hand ('dot com', 'four-one-five') is extra work you don't need. Keep the normal written form, then listen to the result.
Keep each run within 5,000 characters
The Text box takes up to 5,000 characters. Split longer chapters at scene or paragraph boundaries instead of mid-sentence, and put a line break between paragraphs, which is how MiniMax's own text guidance marks paragraph breaks.
Choose Auto or a fixed Emotion on purpose
Auto lets MiniMax pick the delivery from the text, which works well for mixed content. Choose a specific style when a line has to land one way: Calm or Neutral for instructions, Happy for promos, Fearful or Sad for drama. A fixed emotion that doesn't fit the script will clash with it.
Keep Voice ID and Emotion the same across multi-part projects
If a long script is split over several runs, use the same version, Voice ID, Emotion and Speed for every part. The voice preset is what keeps the narrator the same from part to part. Changing settings halfway makes the parts harder to join.
Slow the Speed slightly for dense information
Speed runs from 0.5 to 2.0. For phone numbers, addresses and step-by-step instructions, try about 0.9 so listeners can catch the details. For upbeat promos, 1.05 to 1.15 adds energy. Leave Volume at 1.0 and raise it in small steps only if the file plays too quietly.
Match Sample Rate, Bitrate and Channel to where the audio plays
8000 or 16000 Hz suits phone-line playback. 44100 Hz with a 256000 bitrate suits video, podcasts and narration over music. Keep Channel on Mono for a single voice unless your editing template expects stereo files.
MiniMax Speech 2.6 or Kokoro 1.0: Pick the Right Voice Engine for Your Script
The first five rows show what each generator on this page lets you choose, enter and get back. The last three show what MiniMax and hexgrad publish about the models themselves. Together they help you match an engine to your script's length, languages and casting needs.
| Feature/Spec | MiniMax Speech 2.6 | Kokoro 1.0 |
|---|---|---|
| Versions to pick from | Speech 2.6 Turbo, Speech 2.6 HD | Kokoro 1.0 (single version) |
| Script length per run | Speech 2.6 Turbo and HD: up to 5,000 characters | Up to 7,000 characters |
| Voices to cast from | Speech 2.6 Turbo and HD: 17 preset voices (default Wise Woman) | 40 entries in the Voice list (default Adam - Male) |
| Steering expressiveness | Speech 2.6 Turbo and HD: Emotion, either Auto or one of 9 named styles | Exaggeration slider, 0.25–2.0 (neutral 0.5) |
| Results per run | One audio file | One audio file |
| Developer | MiniMax | hexgrad |
| Model release | October 29, 2025 | January 27, 2025 |
| Languages the model speaks | Speech 2.6 HD and Turbo: 40 languages | 8 languages |
Match the Engine to Your Script and Audience
Long single takes or wide language reach
Kokoro 1.0's text box holds 2,000 more characters per run. A long chapter or lecture therefore splits into fewer pieces, with fewer joins to clean up in the edit. Speech 2.6 HD and Turbo win on reach instead: they cover five times as many languages as Kokoro 1.0, which matters when one campaign has to ship in many markets.
If you mostly produce long single-language narration in a language Kokoro covers, the bigger text box saves editing time. If you localize the same script into many languages, the longer language list decides it.
Casting range or named emotional direction
Kokoro 1.0 has the longer voice list, which helps when a project needs many distinct characters. Speech 2.6 Turbo and HD have fewer voices, but you can name the feeling you want. Pick Fearful for a cliffhanger, Surprised for a twist or Calm for a guided routine, and the delivery leans toward that style. Kokoro 1.0 instead uses a single Exaggeration slider that turns overall expressiveness up or down.
Named emotions suit scripted scenes where every line has a clear intent. A continuous slider suits fine-tuning how animated one narrator sounds. Speech 2.6 also gives you two versions to try the same line on, while Kokoro 1.0 runs as one.
Choose by Script Length, Language Reach and Casting Needs
Use this quick guidance to pick the best option for your workflow.
How to choose: Pick MiniMax Speech 2.6 if your scripts need more languages (Speech 2.6 HD and Turbo cover 40, Kokoro 1.0 covers 8) or if you want two versions, Turbo and HD, to choose between. Pick Kokoro 1.0 if one run has to hold more text (7,000 characters instead of 5,000) or if you want the longer voice list (40 entries instead of 17). Both return one audio file per run, and each has one control for steering expressiveness: named Emotion styles on Speech 2.6, an Exaggeration slider on Kokoro 1.0.
Turn Text into Finished Speech in Six Steps
Six steps take your script from the Text box to a finished audio file.
Choose Speech 2.6 Turbo or Speech 2.6 HD
Open the model list and select the version for this run. Both versions have the same controls on this page, so you can switch between them without rebuilding your settings.
Write or paste your script into Text
Enter up to 5,000 characters of the words you want spoken. If you only have a rough idea, click Enhance Prompt under the box for fuller suggestions. Your text changes only when you pick one.
Cast the read with Voice ID and Emotion
Pick one of the 17 voices (Wise Woman is preselected). Then leave Emotion on Auto or choose a style such as Calm, Happy or Fearful to match the script.
Set pace and level with Speed and Volume
Move the Speed slider between 0.5 and 2.0 to slow down detailed passages or speed up promos. Adjust Volume from its default of 1 only if you need a louder or softer file.
Pick Sample Rate, Bitrate and Channel
Keep the defaults (32000 Hz, 128000 bitrate, Mono) for general use, or change them to suit where the audio will play, such as phone lines, video timelines or podcasts.
Click Generate
Each run returns one audio file. Listen back, adjust the script or settings if a line needs a different feel, and generate again.