Talking Pixel

Talking Pixel

Gives you a pixel character that speaks whatever text you type, using a tiny 1.4M-parameter voice model that runs entirely in your browser with no server and no GPU.

Made by cfahlgren1 on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Talking Pixel is a pixel character that speaks the text you type. The voice comes from PiVo-1.4M, a small text-to-speech model with about 1.4 million parameters, including its vocoder. It runs entirely in your browser, so there is no server and no GPU. The lip movements follow the model's own phoneme timing.

Who it suits: It suits people who want to hear a tiny voice model speak, and developers who want to see how speech synthesis can run fully on the client.

Industry | Who uses it | How they use it | Example
Software | Web developers | Study how a model runs in a page with WebAssembly and a worker | Read the frontend and vocoder source files
Research | Speech researchers | Listen to what a 1.4M-parameter English voice can do | Type a few sentences and judge the quality
Games | Indie creators | Look at a pixel character that talks and moves its mouth | See how lip sync is driven by audio level
Education | Students | Follow a complete browser-side speech pipeline | Compare it with a server-based voice

You can skip it if you need a natural voice in several languages, since this is a tiny English voice that can sound metallic.

Where to use it:

• A web browser: open the Space page on Hugging Face and type your text.
• Windows, macOS or Linux: download the files and open build/index.html in a browser.

Use case | Who | What they get back
Hear text spoken | Anyone curious | Speech from a pixel character, with lip sync
Study the runtime | Developers | A custom ONNX interpreter and a WebAssembly kernel
Compare runtimes | Developers | A benchmark page that compares this runtime with onnxruntime-web

How to set it up: The hosted page needs nothing from you. Downloading a copy needs Python and pip, plus a browser.

• Python, which brings pip, on Windows, macOS or Linux.
• A web browser to open the built page.
• Node, only if you want to develop or rebuild it.

1. Install the huggingface_hub package with pip.
2. Download the Space's files into a folder with the hf command.
3. Move into that folder.
4. Point your browser at the file build/index.html inside that folder.

pip install -U huggingface_hub
hf download cfahlgren1/talking-pixel --repo-type space --local-dir talking-pixel
cd talking-pixel

How to use it: Open the page and type the text you want spoken. The character says it, and playback starts before the whole sentence is synthesized. There is also a benchmark route in the page that compares this runtime with onnxruntime-web.

To work on the project, install the packages and use the scripts below.

npm ci
npm run dev
npm run check
npm run build

Tool | What it does
npm run dev | Runs the project in development mode
npm run check | Checks the project
npm run build | Builds the page and stamps the real download size into it

Example: You open the page and type a sentence that holds a number and an abbreviation. The text frontend normalizes them, the model makes the audio, and the pixel character speaks it while its mouth moves with the sound.

For newcomers and experts: If you are new, open the page and type one short sentence. Listen to it, then try a longer one with numbers in it. Expect a slightly robotic voice.

If you know the field, the text frontend is an exact port of the model's Python text code, with number, money, date and initialism handling and a dictionary lookup. The encoder and decoder run on a small ONNX interpreter whose matrix multiplies use a hand-written WebAssembly SIMD kernel. The vocoder is in the style of FARGAN. A Web Worker does the synthesis and sends audio to an AudioWorklet in 160-sample frames. The page is built with SvelteKit and Svelte 5.

You type some text
The text frontend normalizes it
Encoder and decoder run in a Web Worker
The vocoder streams frames to the AudioWorklet
The pixel character speaks with lip sync

Good to know: The voice is English only and tiny, at 1.4M parameters, so a few sentences can sound metallic. The page needs no onnxruntime-web download.

• License: apache-2 for the PiVo code and weights, BSD-3-Clause for the vocoder derived from Xiph and Opus, and the CMU Pronouncing Dictionary has its own license file
• Cost to run: none named
• Data it reaches: none named, since everything runs in your browser, and the page downloads its model files and a dictionary
• Not for: natural-sounding speech or languages other than English

On Hugging Face, 13 November 2026

• Likes: 2
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/cfahlgren1/talking-pixel
• Files: https://huggingface.co/spaces/cfahlgren1/talking-pixel/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.