About OfflineTTS

OfflineTTS is a free, privacy-first AI voice tool that runs entirely in your browser. It offers both text to speech — powered by Kokoro TTS, Kitten TTS, Piper TTS, and Supertonic TTS — and speech to text — powered by OpenAI Whisper. Speech generation and transcription run in your browser; Kokoro non-English text uses the lightweight phonemization path described below.

Our Mission

We believe useful voice AI should be available without an account, API key, or per-character bill. Running inference in the browser reduces the amount of content a hosted service needs to handle and lets your own hardware do the speech work. It does not mean the website makes no network requests: models must be downloaded, site analytics are described in the Privacy Policy, and non-English Kokoro text uses a narrowly scoped phonemization service. We document those boundaries so you can choose the right engine for your data.

What OfflineTTS Builds

OfflineTTS builds the proprietary browser application that brings several upstream speech models into one practical workflow. That product work includes model selection and delivery, browser caching, long-text chunking, playback, WAV or MP3 export, subtitle export, document reading, and interfaces for choosing an engine, language, voice, and inference backend. The site also maintains first-party setup guidance and explains limitations that affect browser use.

The OfflineTTS team did not train Kokoro, Kitten, Piper, Supertonic, or Whisper. Those are third-party models and runtimes with their own maintainers and licenses. Integrating them into OfflineTTS does not transfer ownership of the upstream projects, and an upstream open-source license does not make the OfflineTTS application open source.

Product Boundaries

OfflineTTS is an interactive browser tool, not a hosted speech API, account-based cloud workspace, or guaranteed production service. The first visit needs a network connection to download application assets and model files. Storage use, speed, memory pressure, and WebGPU availability depend on your browser and device. A model update, cleared browser storage, or a different browser profile can require another download.

Speech and transcripts generated by AI can contain pronunciation, timing, punctuation, or recognition errors. Listen to or read the result before publishing it, and confirm that you have the rights to the source text, recording, voice dataset, and intended use. For sensitive text, choose a workflow whose documented network behavior meets your requirements; English Kokoro, local Supertonic, Kitten, Piper, and Whisper workflows do not use the non-English Kokoro phonemization request after required assets are available.

Who Operates OfflineTTS

OfflineTTS is operated and maintained by the OfflineTTS Project Team. The team is responsible for product releases, site maintenance, privacy documentation, technical articles, and corrections. The OfflineTTS application code is proprietary and its internal development history is not published. Support, licensing, privacy, and correction requests can be sent to contact@offlinetts.com.

Articles use the collective byline OfflineTTS Editorial Team. Our transparency and testing page explains how browser claims are checked, how upstream facts are sourced, which model licenses apply, and how material changes are recorded.

How It Works

When you first visit a tool, it downloads the AI model and caches it in your browser. TTS models range from lightweight Kitten and Piper downloads to Kokoro quality variants and Supertonic's multi-file ONNX model stack. Approximate STT downloads range from ~120MB (Whisper Tiny) to ~590MB (Whisper Small). After that, audio synthesis happens on your device using WebGPU (or WebAssembly as a fallback). Audio synthesis and transcription inputs stay in the browser; the non-English Kokoro exception concerns text sent for phonemization, not an audio upload.

Phonemization Architecture

Why server-side phonemization? Kokoro's browser library only supports English phonemization natively. For the other 7 languages (Japanese, Chinese, Spanish, French, Hindi, Italian, Portuguese), phonemization requires specialized models — misaki for Japanese and Chinese, and espeak-ng for the rest. These models are too large to bundle in a browser (~50MB+), so we run them on a lightweight server.

What does the server do? The phonemization server converts your text into IPA phoneme strings — a few bytes of pronunciation data. It does not generate audio. OfflineTTS configures the service not to retain submitted text as application content; ordinary infrastructure security and request metadata are addressed in the Privacy Policy. Processing time varies with network and server conditions.

Audio synthesis stays local. After receiving the phoneme string, all audio generation happens entirely in your browser via WebGPU/WASM. OfflineTTS does not upload generated audio. English voices are fully offline after model download — they use kokoro-js's built-in phonemizer and never contact the server. Supertonic uses language-tagged text and runs its synthesis path locally after the model files are downloaded.

Technology

• Kokoro TTS — 82M parameter open-source TTS model (Apache 2.0)
• Kitten TTS — lightweight TTS with expression-based voices
• Piper TTS — fast CPU-optimized TTS with 25+ voices
• Supertonic TTS — multilingual on-device TTS with 31 languages and 10 built-in voices
• Whisper STT — OpenAI's speech recognition model, 99 languages
• browser-whisper — production-oriented Whisper wrapper with Web Workers and streaming
• ONNX Runtime Web — browser-based ML inference via WebGPU/WASM
• Astro — fast, SEO-optimized static site framework
• Cloudflare Pages — edge CDN for global delivery

Upstream Models and Runtimes

OfflineTTS builds on browser-friendly AI projects including Kokoro, Kitten, Piper, Supertonic, and Whisper. Their upstream licenses apply to those models and runtimes; they do not make the OfflineTTS application itself open source. OfflineTTS does not claim to have trained these upstream models. Source links, recorded licenses, integration qualifications, and product-level ownership are documented on the transparency page.