Generate speech on your device: Keep your text local
Audio generation uses a local speech model to turn your text into a WAV on this device. The text you enter, and any reference audio you provide, stay on your device during generation. Once the app, runtime and model files are available, no inference-server connection is needed; downloading missing files or updates still uses the network.
First, open /#/audio-generation to start in the audio workspace.
1. Open Audio generation and prepare a model
The link above opens Audio generation directly. You can also open More Actions in the sidebar and choose Audio generation. For a first model, find Qwen3-TTS 1.7B Base in the open Model catalog and select Download. This download includes the required audio mmproj. When it finishes, select Use this model. If you already have a suitable model, select it directly. For other models, open Models and runtime to import a matching backbone GGUF and audio mmproj together or obtain the pair from a repository. Local-only cannot fetch from an external repository; obtain both files separately and import them there.
2. Enter text and choose its language
Choose an Audio model, enter Text to speak, and set Language for the speech you want. The initial choice follows Naidan’s interface language; it does not detect the language of your text. Qwen’s Model default means English, while Pocket TTS language is determined by its weights. Qwen3-TTS Base can generate without reference audio; Pocket TTS requires a reference voice.
3. Generate and follow progress
Select Generate audio. Progress appears on the page, and a completed result is added under Generated audio. During a longer run, Get audio so far adds an intermediate WAV while generation continues; Finish generation and keep audio ends with the audio accumulated so far. Cancel and discard throws away the uncaptured part.
4. Play the result and save its WAV
Use the result player when you choose, then select Save WAV to keep a file. Completed results and their settings remain in this page’s memory only. Leaving or reloading clears them, and playback never starts automatically. Save anything you need before leaving.
Add a reference voice when needed
Pocket TTS needs a reference voice; Qwen3-TTS Base can use one but does not require it. The reference section accepts a local file, or a microphone recording when your browser and permissions allow it. A page-local reference library can keep several clips selected, up to 30 seconds total selected audio. The reference clips are not copied into generated-result history. The model must support the reference voice path; selection alone does not guarantee a particular voice or quality.
Chat Read aloud is a separate feature
Read aloud on a chat answer uses the browser’s Web Speech API. It does not create a WAV in this workspace. Where that voice is processed depends on the browser, operating system and selected voice, so its offline behavior is different from this local model generation.