Turn Scripts Into Natural Narration With Kokoro 1.0
Last verified: November 3, 2026
Kokoro 1.0 fits speech quality that its creator rates as comparable to much larger text-to-speech models into just 82 million parameters, and it runs significantly faster . A typed script comes back as natural-sounding narration for voiceovers, lessons or audiobook chapters. It is the v1.0 release of hexgrad's open-weight TTS model. It was published on January 27, 2025, and is distributed on Hugging Face as Kokoro-82M . The design builds on StyleTTS 2 and the ISTFTNet vocoder. Text reaches the model through misaki, hexgrad's own grapheme-to-phoneme library .
Version 1.0 is a big step up from the v0.19 release of December 25, 2024. Training data grew from under 100 hours to a few hundred hours, and coverage went from 1 language with 10 voices to 8 languages with 54 voices . The voices cover American and British English, Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian and Brazilian Portuguese . That gives one compact model enough voices for stories with several characters and for narration in other languages.
Kokoro was trained only on permissive, non-copyrighted audio paired with IPA phoneme labels . That matters to teams who need to check where a voice model's training audio came from. On this page, the model runs as a text-to-speech generator with a 40-voice menu and sliders for expressiveness and variation.
Speech Generator Limits at a Glance
The script, voice and delivery settings that shape every run, plus the size of the model behind them.
Script length per run
Up to 7,000 characters of text
Voice menu
40 speakers; opens on Adam - Male
Exaggeration (expressiveness)
0.25–2.0 in 0.01 steps, neutral at 0.5
Temperature (variation)
0.05–5.0, default 0.8
Model size
82 million parameters, open weights
Output per run
One audio file
Six Checks Before You Press Generate
Avoid the most common Kokoro problems: a voice in the wrong language, a weak speaker, or unstable delivery.
Try a stronger voice than the default
Voice opens on Adam - Male, but hexgrad grades Kokoro's Adam voice F+, the bottom of its scale, because its reference audio rates only D for quality . For lead narration, try Bella (A-), Nicole or Emma (both B-) first. Among the American male voices, Fenrir and Michael grade C+.
Write the script in the voice's own language
Kokoro links each voice to a pronunciation pipeline for one language, and its creator's code says the language must match the voice . Give French text to Siwis, Italian to Sara or Nicola, and Mandarin to the Xiao- and Yun- voices. The list has two entries named Alpha. The first sits with Beta, Omega and Psi, which are Kokoro's Hindi speakers. The second sits with Gongitsune and Kumo, its Japanese speakers.
Keep Exaggeration close to 0.5 for narration
0.5 is the neutral point, and the control's own note warns that extreme values may be unstable. Stay near neutral for lessons and audiobooks. When a character line needs more energy, raise it in small steps and listen before going further.
Lower Temperature for an even long-form read
Temperature starts at 0.8, and higher values make the delivery more varied. If the narration should sound the same from the first paragraph to the last, stay at or below the default. Save higher values for trying different readings of a short line.
Fit long scripts to the 7,000-character box
Split books and courses at scene or section breaks, never mid-sentence. Run every part with the same Voice, Exaggeration and Temperature so the same speaker carries through. Clear punctuation and paragraph breaks give the voice natural places to pause.
Kokoro 1.0 or MiniMax Speech 2.6: Choose the Engine for Your Script
Kokoro 1.0 and MiniMax Speech 2.6 both turn typed scripts into speech on Vidofy. The first five rows show what each generator here lets you enter and adjust. The last two give language coverage and release dates as each model's creator publishes them.
| Feature/Spec | Kokoro 1.0 | MiniMax Speech 2.6 |
|---|---|---|
| Versions you can run | Kokoro 1.0, text to speech | Speech 2.6 Turbo or Speech 2.6 HD, both text to speech |
| Script length per run | Up to 7,000 characters | Up to 5,000 characters (Turbo and HD) |
| Voices to choose from | 40 voices in Voice, starting on Adam - Male | 17 voices in Voice ID, starting on Wise Woman (Turbo and HD) |
| How delivery is steered | Exaggeration slider, 0.25 to 2.0, neutral at 0.5 | Emotion menu: Auto or one of nine named styles (Happy, Sad, Angry, Fearful, Disgusted, Surprised, Calm, Fluent, Neutral), default Auto (Turbo and HD) |
| Audio files per run | One audio file | One audio file (Turbo and HD) |
| Language coverage | v1.0: 8 languages, 54 voices | More than 40 languages |
| Release date | January 27, 2025 | October 29, 2025 |
Matching Each Engine to the Job
Script length and voice range
Kokoro 1.0's text box holds 40 percent more text per run than Speech 2.6's. An audiobook section or a full lesson needs fewer splits, and fewer joins later in editing. Its voice menu is also more than twice as long. It mixes American and British English speakers with French, Hindi, Italian, Japanese and Mandarin voices, which helps when one project needs several characters or several markets. Speech 2.6, for its part, lets you choose between two versions here: Turbo and HD.
How each one shapes delivery
The two generators control the performance in different ways. Kokoro's Exaggeration slider works like an intensity dial: starting from neutral, you set how animated the read is in fine steps. Speech 2.6's Emotion menu works by name: you pick a mood such as Calm or Surprised, or leave it on Auto. Either way, write the script first, then use that engine's control to match the tone of the text, and listen after each change.
Language reach of each model
Kokoro v1.0 spreads its 54 voices over eight languages, while MiniMax says Speech 2.6 supports more than 40. If your script is in English, French, Hindi, Italian, Japanese or Mandarin, Kokoro's Voice menu here has speakers built for that language. For a language outside Kokoro v1.0's eight, Speech 2.6 is the model designed to cover it.
Choosing Between Kokoro 1.0 and MiniMax Speech 2.6 for Your Next Script
Use this quick guidance to pick the best option for your workflow.
How to choose: Choose Kokoro 1.0 when a script runs long, since its Text box holds 7,000 characters against 5,000 on Speech 2.6 Turbo and HD. It is also the pick when you want more voices to choose from: 40 against 17. Choose MiniMax Speech 2.6 when the script is in a language outside Kokoro v1.0's eight, since Speech 2.6 covers more than 40. It is also the pick when you want to choose between two versions, Turbo and HD. Both return one audio file per run.
From Script to Spoken Audio in Four Moves
Four steps from an empty text box to a finished voice track.
Write or paste your script in Text
Enter up to 7,000 characters. Everything in the box is read aloud, so remove notes and stage directions. If you only have a rough idea, the Enhance Prompt button under the box suggests fuller wording. Your text changes only when you pick a suggestion.
Choose a speaker from Voice
Open Voice and choose one of 40 speakers; the menu starts on Adam - Male. Pick a voice that matches the language your script is written in.
Set Exaggeration, Temperature and Seed
Leave Exaggeration at 0.5 for a neutral read, or raise it for livelier delivery. Set Temperature lower for a steadier read or higher for more variation. Seed takes a number if you want to enter one.
Press Generate and listen back
Click Generate to create one audio file from your settings. Listen to the result, then adjust any slider before you run it again.