TTS.Rocks

Turn text into speech in your browser. Choose local AI voices, clone a reference voice locally, or use a cloud provider with your own API key.

Open TTS.Rocks

1. Enter text and choose an engine.
2. Open Voice to choose a voice and language. Adjust speed and playback timing under Advanced.
3. Click Generate Speech. Local models download on first use and are cached when browser storage is available.
4. Play, pause, seek, or download the recording.

Use the shortcuts above the studio for Voiceover, Read a document, Clone a voice, or Summarize. In the Voice tab, Hear buttons play six prerecorded Kokoro and Kitten starter voices without downloading a model. Selecting a sample chooses that voice for your next recording.

Save projects and download files

Open Projects above the editor to name and save a script, voice settings and optionally its last completed audio. Reopen it later in the same browser to continue editing or download the saved recording. Save project updates the opened project; Save as new project makes another copy. Projects exclude API keys, reference recordings and music files; select those separately when reopening. Unfinished document sections stay in the current tab, so finish generation before saving its audio.

Expand Save & export to set a recording filename or download a ZIP containing the last completed audio, its script, timing JSON and any VTT captions. Editing the script does not change existing audio: generate again to include your changes. ZIP bundles and saved project audio support recordings up to 128 MB; for larger recordings, download the audio separately and save the project with Include the last completed audio unchecked. Browser storage can be cleared or evicted, so keep downloaded copies of important work.

Read a document

Choose a PDF, TXT, Markdown, Word (DOCX) or ODT file using Upload file. Review the extracted text, select voices, then click Generate Episode. Use Single voice for narration or Alternate narrators to share the reading. An existing A:/B: script can also use two voices. The document is read as written; edit or summarize it first if you want a shorter discussion.

Files are read locally. Document mode supports up to 100,000 characters and 25 MB per file with Kokoro, Kitten 0.8, Supertonic or Chatterbox; Chatterbox uses one voice. Choose Long text for longer pasted scripts. For scanned PDFs, run OCR first; for older .doc files, export as DOCX or text. Review PDF reading order and any pages reported as missing text.

Download the complete WAV, captions (VTT), or individual sections. Stop keeps completed sections in the current tab; Generate Episode resumes if the text and voice settings are unchanged. Keep the tab open until downloads are saved. Add music to the exported episode in your audio or video editor.

Edit text locally

Choose Local AI to fix spelling and grammar or summarize a short passage. Select text first to edit just that passage (up to 3,000 characters). Review the editable draft, then choose Use this text.

Qwen3-0.6B uses Apache-2.0 and runs on your device with WebGPU. First use downloads about 0.6–1 GB from Hugging Face into the browser cache; no API key or server is needed. Check names, numbers and meaning before applying a draft. Model & download lets you clear its cached files.

Choose an engine

Engine	Runs in	Best fit
Kokoro	Browser CPU or WebGPU	28 US/British English voices plus eight blends; compact Q8 or full-quality FP32
Chatterbox	Browser CPU or WebGPU	Expressive English speech and reference-voice cloning; roughly 1.5 GB download
Kitten 0.8	Browser CPU	Eight English voices; Nano (25 MB), Micro (41 MB), or Mini (80 MB)
Supertonic 3	Browser CPU or WebGPU	Ten voices, 31 languages, adjustable generation steps
Pocket TTS	Browser CPU	Built-in voices and local reference-voice cloning; six language bundles
MusicGen Small	Browser CPU	Short instrumental clips; non-commercial model, roughly 660 MB download
Piper	Browser CPU	Bundled US and British English voice models
eSpeak	Browser CPU	Lightweight multilingual synthesized speech
Kitten 0.1	Browser CPU	Original Kitten model and voices
Browser Native	Browser/operating system	Immediate speech; no model download or audio export
ElevenLabs	Provider API	Flash v2.5, expressive v3/v4, and account voices
OpenAI	Provider API	GPT-4o mini TTS with voice directions, TTS-1 and TTS-1 HD
Google Cloud	Provider API	Standard, WaveNet, Neural2, Studio and available account voices

Kokoro's automatic mode uses FP32 on WebGPU and Q8 on CPU. Smaller models reduce downloads but are not necessarily faster on every GPU. Choose CPU if GPU initialization fails. Supertonic defaults to CPU in automatic mode. Sizes exclude runtime and voice files.

Kokoro blends combine the two named voices equally: for example, Heart + Bella blend. The original voices remain available. Kokoro model and voice assets use Apache-2.0; Chatterbox model and default reference use MIT.

Choose Chatterbox for expression controls or to match a reference voice. It needs a desktop with room for the larger model; preparation and CPU generation can take several minutes. Try WebGPU for faster generation. Choose Subtle, Natural, Expressive or Dramatic under Advanced > Expression, and adjust playback speed in the player.

Choose Advanced > Start playback > As audio arrives to hear the first chunk while the rest is generated. This is selected for new users; saved playback preferences are preserved. Slower devices may pause between chunks. Choose complete-recording playback for uninterrupted listening. The download contains every generated chunk.

The display in Advanced reports the actual CPU or GPU after the model loads and updates generation speed as chunks finish. 2x real time means ten seconds of audio generated in five seconds. This rate excludes model loading and added pauses; the first-chunk time includes loading. Playback has a short silent lead-in to give the audio output time to start; exported audio and caption timestamps are unchanged.

During early playback, use the live pause/resume and volume controls. Generation continues while playback is paused. To remove a local engine's cached downloads, select it and use Advanced > Clear downloads for this engine; it downloads again on next use.

In Advanced, choose Mobile / low latency for Kitten Nano and early playback, or Quality narration for Kokoro. Advanced > Download / prepare model loads supported models before recording. Keep the tab open to reuse the loaded model. Check this device reports WebGPU and FP16 availability; Advanced also offers lower-power and performance GPU preferences. The browser decides which hardware is available.

Generate narration with an AI assistant

Open Use with an AI assistant to copy Markdown instructions, your current voice settings, or JavaScript and CLI examples into your coding agent. Agents can navigate the labelled controls or call window.ttsRocks inside the page.

The CLI helper uses headless Chromium to save WAV audio, VTT captions and JSON timing metadata, including named batches for video narration. Use the live website or a local copy with --local. Kokoro, Kitten 0.8, Supertonic and Chatterbox support direct calls. See the automation guide for Windows, macOS and Linux setup.

Clone a voice locally

Select Chatterbox for English expression controls, or Pocket TTS for a smaller model with English, French, German, Italian, Portuguese and Spanish bundles. Upload a clear recording of your voice or one you have permission to use. A 5-10 second sample works well; the first 10 seconds are used. Leave the reference empty to use a built-in voice.

The reference is encoded in a browser worker and is not uploaded. To reuse it after closing the page, expand Save or reuse a reference voice, give it a name and choose Save reference on this device. Select it from Saved reference and click Use reference on your next visit. Delete saved reference removes the stored copy; clearing site data also removes it. Pocket English needs approximately 125 MB of model downloads, plus approximately 21 MB for the cloning encoder. Other language bundles can be larger. Chatterbox downloads approximately 1.5 GB, including its cloning encoder. Use a desktop browser if your device runs out of memory.

The same reference is reused for subsequent recordings. Clear reference removes the selected file and releases its encoded voice from memory.

Make a two-speaker podcast

Choose Two speakers (A:/B:) under Text Input and use Kokoro, Kitten 0.8 or Supertonic. Select Speaker A in the main Voice selector and Speaker B in the podcast controls. Write each turn with a label:

A: Welcome to our show. What are we talking about today?
B: How to make a podcast entirely in the browser.
A: Let's get started.

Set the pause between speakers, generate, then download the complete WAV and speaker captions (.vtt). Captions follow speaker turns rather than individual words. Labels are not spoken.

Expand Music bed and intro / outro to add a local track or the last MusicGen clip generated in this page. The music loops with fades and becomes quieter under speech. Music beds use complete-recording playback, and captions include the intro offset. Files remain local and must be selected again after reloading the page.

Create an instrumental clip

Choose MusicGen, describe the instruments, mood and rhythm, and select a 3–15 second clip. Generation runs locally on CPU and can take minutes; use a desktop with room for the model. Download the WAV or select it as a podcast music bed.

MusicGen uses CC BY-NC 4.0 (non-commercial) model weights. Review the model license before choosing it for a project. For lyric-based songs and larger local music models, see ACE-Step 1.5, which runs in a separate native installation.

Cloud voices and translation

For ElevenLabs, OpenAI or Google Cloud, enter your own API key. Requests go directly from your browser to that provider. Keys are saved in this browser's local storage; do not use a shared browser profile for private keys. Provider charges and account limits apply.

Load my available voices retrieves current ElevenLabs or Google voices, including voices available to your account. For GPT-4o mini TTS, add a direction such as "warm, calm narration" in the Voice tab. Older OpenAI models use their own supported voice list.

Voice choices are remembered per engine and language. Loaded account voice lists stay available while switching engines in the same page session.

GPT-4o mini TTS offers warm, calm, upbeat and newsreader direction presets. Edit the direction to suit your script. Local speech engines use their own voice and speed controls.

Eleven v4 accepts up to 2000 characters per recording here. Use the player's speed control for Eleven v3/v4; their synthesis speed controls are unavailable. Choose Flash v2.5 or another existing model for Caption.Ninja links.

On compatible desktop Chrome installations, the text toolbar offers local summarization, language detection, rewriting and translation. Choose the target language in Voice, then use Translate. Browser language models may need an initial download. Review edited text before generating speech.

Run your own copy

Install Node.js 22.12 or newer. Clone the repository, or download and extract the source ZIP. No build or npm installation is needed to open the local studio:

git clone https://github.com/steveseguin/tts.rocks.git
cd tts.rocks
node local-server.mjs

Open http://127.0.0.1:8844/. The server listens only on this computer; stop it with Ctrl+C. Use --port 8845 to choose another port. You can also serve the folder with your own HTTPS or localhost server. Use a current browser with WebAssembly and Web Workers. WebGPU requires a compatible browser, graphics driver and secure context. Models are fetched from Hugging Face, and the neural WASM runtime from jsDelivr. Local synthesis keeps your text and reference audio on your device; the source ZIP does not bundle model weights or guarantee offline use.

For command-line generation, run npm run setup:cli once in the extracted folder, create narration.txt, then run:

node tts-rocks.mjs --local --input narration.txt --output narration.wav

Local mode starts and stops its server automatically. See the automation guide for voice selection, batches and JSON requests.

If your browser cannot preview audio, download the recording and play it in another app. Reference cloning requires a browser that can decode audio locally.

For the optional Vite workflow, use Node.js 22.12 or newer:

npm install
npm run dev
npm run build
npm run preview

The build writes a static site to build/, including the standalone TTS library and its assets. Serve the whole output folder.

Caption.Ninja and embedding

Expand Use with Caption.Ninja to create speech-input, manual-input and output links for one room. Kokoro, Piper, eSpeak, legacy Kitten, browser speech and cloud engines can be passed to existing overlays. Kitten 0.8, Supertonic and Pocket are studio options. Generated cloud-provider links include API keys: share them only with trusted recipients.

For a simple browser-speech integration:

<script src="https://tts.rocks/tts.js"></script>
<button onclick="TTS.TTSProvider = 'system'; TTS.speak('Hello from TTS.Rocks', true)">
  Speak
</button>

The tts.js library and the studio UI are separate entry points. Embed the complete studio with an iframe when you need its model controls, cloning and waveform player.

More voice model projects

These projects offer native runtimes, hosted demos or community browser ports. Use their linked setup instructions; they are not additional engines in the TTS.Rocks selector.

• Chatterbox: Nano, Turbo and multilingual models, cloning and expressive speech. Community ONNX exports offer a separate browser integration path.
• Qwen3-TTS: multilingual cloning, custom voices and voice design, with native model runtimes.
• NeuTTS: Air and Nano cloning models, plus 2E expression controls. Choose the appropriate language, codec and model license.
• KittenTTS 2: a separate speech-language model with cloning and expression controls. Its native runtime differs from the lightweight Kitten 0.8 ONNX models used here.

Licenses and support

TTS.Rocks application code is MIT-licensed. Third-party code, model weights and voice assets retain their own licenses. Supertonic 3 weights use BigScience Open RAIL-M; Pocket model/voice assets have separate attribution requirements. See neural component notices and each bundled library's license before redistributing assets.

Report an issue. Created by Steve Seguin.