Kokoro-82M Explained: Why Local TTS Uses It On-Device

When a text-to-speech app reads a page aloud, the controls you can see are only part of the experience. The model behind them determines whether a voice sounds natural, keeps a steady rhythm, pronounces a sentence clearly, and responds quickly enough to feel like a reader rather than a rendering job.

Local TTS uses Kokoro-82M for its American and British English neural voices. The model is small enough to run on an iPhone or iPad, yet capable enough to make long-form listening feel natural. Best of all, the text and generated speech stay on your device.

You do not need Python, a terminal, model files, or an API key. Paste text, import a document, or scan a printed page in Local TTS and press Play.

Official image and technical sources: The cover is Hugging Face’s official social preview for the hexgrad/Kokoro-82M model repository, used unchanged—not an AI-generated illustration. Model architecture, release, training, and license details below are cited to that model card; voice counts and quality limits are cited to the official VOICES.md.

The one-sentence answer

Kokoro-82M is an 82-million-parameter, open-weight text-to-speech model that delivers natural 24 kHz speech with far less compute than many larger voice models—making it a strong fit for private, on-device reading.

That balance matters. A model can sound wonderful on a server with a powerful GPU and still be the wrong choice for a phone. Local TTS needs natural speech, practical memory use, responsive playback, and no trip to a cloud API. Kokoro makes that combination possible for English.

What is Kokoro-82M?

Kokoro-82M is a neural text-to-speech model created by hexgrad. “82M” refers to its roughly 82 million parameters. The official Kokoro model card describes it as an open-weight model designed to offer quality comparable to larger systems while remaining faster and more cost-efficient.

Version 1.0 was published on January 27, 2025. According to the model card, it was trained on a few hundred hours of permissive, public-domain, and synthetic audio. Its weights use the Apache 2.0 license.

Kokoro draws from StyleTTS 2, a system designed to model expressive speaking styles, and ISTFTNet, a lightweight neural vocoder that efficiently turns model output into an audio waveform. The public Kokoro release is not the entire StyleTTS 2 training system: its model card describes it as decoder-only, without the diffusion component or a released encoder.

In plain language, Kokoro packages the part needed to turn phonetic text and a selected voice style into audio, without carrying the weight of a much larger general-purpose AI model.

Why 82 million parameters matter on a phone

Parameter count is not a score. A larger model may capture more nuance, but it also usually needs more memory, computation, energy, and time. Those costs become obvious when an app needs to read a chapter instead of generating one short demo sentence.

Kokoro’s compact size gives Local TTS room to build a complete reading experience around the model:

• On-device generation: Speech is produced locally instead of waiting for a remote server.
• Responsive reading: The app can prepare upcoming sentences while the current sentence plays.
• Long-form practicality: PDFs, EPUBs, and articles can be read without paying per character to a TTS API.
• Offline availability: The voice still works on a plane, during a commute, or anywhere the connection is unreliable.
• Privacy by architecture: Your pasted text and imported documents do not have to be uploaded for speech generation.

This is why “small” is a product advantage here. The goal is not to win a parameter-count contest. It is to make high-quality speech available whenever you tap Play.

Why Local TTS uses Kokoro-82M

Using a model inside a real reading app takes more than placing weights in the app bundle. The text has to be normalized, divided at sensible boundaries, converted to speech ahead of playback, cached, highlighted, and kept playing when the screen locks.

1. Natural American and British English voices

Kokoro v1.0 publishes 20 American English voices and 8 British English voices in its official voice catalog. Local TTS includes all 28 of these English voices in its full catalog, with different accents, genders, and vocal character.

Three high-quality English voices are available in the free version, so you can hear the model on your own text before deciding whether you need Pro.

2. The model is bundled, not called through an API

Local TTS carries the speech model with the app. When you enter text, the app turns it into speech on your iPhone or iPad. There is no TTS provider receiving the document, no server queue, and no account required.

That is especially useful for material you may not want to paste into an online service: an unpublished manuscript, work document, personal note, or study material. If you use iCloud sync, your notes can sync through your own iCloud account; speech generation itself still happens on the device.

3. Two optimized paths keep listening practical

Local TTS uses Apple’s MLX and Metal technologies for responsive foreground generation. It also uses an optimized Core ML path so speech can continue to be prepared during background playback and audio export without depending on a cloud fallback.

You do not have to manage either path. The app chooses the appropriate engine while preserving the same voice and reading controls.

4. A voice model becomes a complete reader

Kokoro generates speech. Local TTS turns that capability into something you can use every day:

• Paste an article or write a note.
• Import PDF, EPUB, Word, TXT, and Markdown files.
• Scan a printed page with on-device OCR.
• Follow word-by-word highlighting during playback.
• Change the voice and reading speed for each note.
• Keep listening from the Lock Screen or Control Center.
• Export speech as an M4A file with Pro.

Download Local TTS free on the App Store to try Kokoro-82M without installing a model or opening a terminal.

Does Kokoro provide all 31 Local TTS languages?

No—and this distinction is important. Kokoro powers Local TTS’s American and British English voice catalog. The app’s 31-language voice is powered by Supertonic 3, a separate multilingual on-device TTS engine.

Read our Supertonic 3 guide to see how one Local TTS voice handles all 31 supported languages offline.

The public Kokoro v1.0 release includes voices across eight languages, but the project’s own voice notes warn that non-English support can be thin because pronunciation tooling and training data vary by language. Local TTS currently uses Kokoro where its mobile implementation is strongest: English. Supertonic 3 handles Korean, Japanese, Spanish, French, German, and the rest of the app’s supported languages.

Both engines run locally. Choosing another language does not switch you to a cloud service.

What affects real-world voice quality?

The same voice will not sound equally natural on every piece of text. These details make a noticeable difference:

1. Punctuation: Commas, periods, and paragraph breaks give the model useful phrasing cues.
2. Names and specialist terms: Uncommon names, acronyms, and brand names may need phonetic spelling or a small text edit.
3. Sentence length: Extremely short fragments can sound abrupt, while very long run-on sentences may rush.
4. Numbers and symbols: Dates, Roman numerals, formulas, and unusual symbols can be ambiguous when spoken.
5. Voice choice: A voice that works well for a novel may feel less suitable for proofreading or technical notes.

Local TTS normalizes common edge cases and splits long text into practical synthesis units. For the best result, keep the source text punctuated and preview a few voices with the material you actually plan to hear.

Kokoro is impressive, not infallible

The official voice catalog notes that voice quality varies with the amount and quality of training data. It also says voices tend to work best in a middle range of roughly 100–200 tokens; very short utterances may be weaker, and very long ones may rush. Local TTS reduces these extremes through sentence-aware preparation and chunking, but it cannot make every pronunciation perfect.

Kokoro is also an open-weight model, not a complete open-source training release. The downloadable weights are Apache-licensed, while the model card says the encoder and full training system are not released.

For pronunciation-critical work, preview the result and check names, abbreviations, and technical language before exporting or publishing the audio. Any shared audio should also be identified as synthetically generated speech.

Frequently asked questions

Does Local TTS really use Kokoro-82M?

Yes. Kokoro-82M v1.0 powers the American and British English neural voices in Local TTS. The model weights and voice data are bundled for on-device use.

Does it work without internet?

Yes. Once Local TTS is installed, text-to-speech generation works offline. No internet connection is needed to turn your text into audio.

How many Kokoro voices are in Local TTS?

Local TTS’s full catalog includes 28 Kokoro English voices: 20 American and 8 British. The free version includes three selected English voices, plus the app’s multilingual voice.

Does Local TTS upload my PDF or text to generate speech?

No. Imported files, pasted text, OCR, and speech generation are processed on the device. There is no account, no advertising, and no tracking.

Do I need to download Kokoro separately?

No. You do not need Hugging Face, Python, a command line, or an API key. Local TTS packages the model and reading interface together.

Is Kokoro used for Korean and the other supported languages?

Not in Local TTS. Kokoro is used for the English catalog; Supertonic 3 provides the app’s 31-language on-device voice.

Hear Kokoro-82M on your own text

Kokoro-82M matters because it makes natural neural speech practical beyond a demo page or cloud API. Local TTS brings that model into an offline reader with document import, OCR, highlighting, background playback, voice controls, and audio export.

Paste a paragraph, import the PDF you have been meaning to read, or photograph a printed page. Then choose an English voice and listen—your text stays on your device from start to finish.

Download Local TTS free on the App Store and hear Kokoro-82M on your iPhone or iPad.