Parakeet.wgsl

Parakeet.wgsl is a browser-based speech-to-text tool that runs the NVIDIA Parakeet TDT 0.6B v2 model entirely on your device. You upload an audio file or select a sample, and transcription happens locally through WebGPU and WebAssembly. Nothing leaves your browser. The model downloads on first use, caches automatically, and subsequent sessions start faster because the weights are already stored. The project comes from Hamza Q. at Narcotic Software and is available on GitHub and npm for anyone who wants to inspect, fork, or extend it. Released in 2026, it represents one of the first serious demonstrations of running a production-grade speech model entirely in the browser without server assistance.

The problem it solves is straightforward. Most transcription services require uploading audio to a server, which means your recordings pass through someone else's infrastructure. For sensitive meetings, legal depositions, medical dictation, therapy sessions, journalistic interviews, HR investigations, or anything you would rather keep private, that creates friction. Even if you trust the service, compliance requirements may prohibit sending certain audio off-device. Cloud transcription APIs also tend to charge by the minute, so high-volume users face ongoing bills that add up quickly. Parakeet.wgsl skips both issues entirely. The compute happens on your GPU, in your browser, and the audio never touches the network. There is no account, no usage tracking, no API key, and no per-minute billing.

Under the hood, the project packages the Parakeet model for WebGPU execution. WebGPU is a newer browser API that provides direct access to the GPU, bypassing the older WebGL abstraction and enabling general-purpose compute workloads. That makes running a 600-million-parameter speech recognition model feasible on consumer hardware, something that would have been impractical in the browser just a few years ago. WebAssembly handles the parts that benefit from compiled code, and together the two technologies let the browser act as a local inference engine. The interface shows performance metrics including audio length, wall time, and transcription speed in real-time factor, so you can see exactly how your machine handles the workload. Those metrics help you gauge whether your hardware is up to the task before committing to a long transcription job.

English audio is the only supported language. The model was trained on English speech, so other languages will produce poor results or fail outright. That is a hard limit of the underlying Parakeet model from NVIDIA, not something the wrapper can fix. If you need multilingual support, this is not the tool for you. Format support varies depending on your browser's media decoders, and the project notes that some files can be transcribed even if the browser's audio player cannot play them back. In practice, common formats like MP3, WAV, M4A, and FLAC work, but edge cases may require conversion before upload.

The tool fits several audiences. Developers exploring on-device AI can use it as a reference implementation for WebGPU inference. Privacy-focused users who want local transcription without trusting a cloud service can transcribe meetings, voice memos, or dictation without any data leaving their machine. Researchers working with sensitive audio, including fields like medicine, law, and journalism, get a workflow that satisfies strict data handling requirements. And anyone curious about what WebGPU can do with a real machine learning model can see a concrete, working example. Developers can also pull the underlying npm package and integrate the same model into their own web applications, which opens doors for local-first apps that need transcription without server calls or per-minute API costs.

Where it differs from cloud transcription APIs is the complete absence of a server. You open the page, load the model, and transcribe. The tradeoff is that performance depends entirely on your hardware. A recent discrete GPU handles the 600-million-parameter model well, producing transcriptions faster than real-time playback on capable machines. Older or integrated graphics will struggle with speed, and underpowered devices may fail to load the model entirely. Browser support also matters. Chrome and Edge have the broadest WebGPU support as of now, while Safari and Firefox lag behind or require flags. If your browser does not support WebGPU, the tool will not work at all.

Access is free. There is no paid tier, no premium features locked behind a subscription, and no rate limits. The project is open source under a permissive license, so you can audit the code to verify that audio stays local and nothing phones home. For anyone who wants transcription without sending files anywhere, this is one of the few options that actually delivers on that promise without compromise. The first-run model download is large, around 600 megabytes for the weights, so initial load takes longer on slow connections, but caching means you only pay that cost once per browser. Subsequent sessions start in seconds.