Show HN: A 40MB TTS model hand-ported to run in the browser (TS and WGSL)

TTSLab launched as an MIT-licensed, open-source project that runs text-to-speech and speech-to-text models entirely in the browser via WebGPU and WASM, with no server-side processing, API keys or data collection. The site's featured models include Kokoro 82M (80MB, 82M parameters, 21 voices), Piper EN-US Lessac Medium (75MB, 50M parameters, part of a Piper project spanning 904+ voices), Microsoft's SpeechT5 (60MB, 60M parameters) and Moonshine Tiny fine-tuned for Korean (35MB, 27M parameters). Each model is downloaded once and cached locally, which TTSLab says eliminates the privacy and compliance risks of sending sensitive text to cloud TTS services and removes network round-trip latency for voice agents and live captioning.

Test Speech AI In Your Browser Test TTS & STT models in your browser. No server. No data collection. Powered by WebGPU. “” Try it yourself https://ttslab.dev/models/kokoro-82m Stay in the loop Subscribe for the latest news on voice AI and text-to-speech — new models, features, and benchmarks. Featured Models View all https://ttslab.dev/models TTS080 MB82M params21 voicesWebGPUWASMTry it https://ttslab.dev/models/kokoro-82m Kokoro 82M High-quality TTS with 21 voices. Best quality-to-size ratio for browser inference. Supports streaming generation. TTS075 MB50M params1 voiceWASMTry it https://ttslab.dev/models/piper-en-us-lessac-medium Piper EN-US Lessac Medium Fast, lightweight TTS using the VITS architecture. Part of the Piper project with 904+ voices across languages. TTS060 MB60M params1 voiceWebGPUWASMTry it https://ttslab.dev/models/speecht5 SpeechT5 Microsoft's unified-modal SpeechT5 model. Lightweight and CPU-sufficient for basic TTS needs. STT035 MB27M paramsWebGPUWASMTry it https://ttslab.dev/models/moonshine-tiny-ko Moonshine Tiny Korean Moonshine Tiny fine-tuned for Korean. Specialized single-language model for higher accuracy than multilingual alternatives. Voice Agent Talk to an AI assistant running entirely in your browser. Try it https://ttslab.dev/voice-agent Text-to-Speech in the Browser TTSLab lets you run text-to-speech models directly in your browser using WebGPU and WASM. No server-side processing, no API keys, no queue times. Models like Kokoro 82M, SpeechT5, and Piper generate natural-sounding speech in real time, right on your device. Each model is downloaded once and cached locally for instant reuse. Whether you are a developer evaluating TTS options, a researcher benchmarking model quality, or a product team comparing voices, TTSLab provides a fast, standardized environment. Why On-Device Speech AI Matters Cloud-based TTS services require sending your text to external servers. For sensitive content — medical notes, legal documents, personal messages — that creates privacy and compliance risks. On-device inference eliminates this entirely: your text and audio never leave your browser. WebGPU-accelerated inference also removes latency from network round trips, making real-time applications like voice agents and live captioning practical without a backend. How It Works Pick a Model Browse our directory of TTS and STT models, or compare two side by side. Downloads Once The model weights are downloaded to your browser and cached locally for instant reuse. Runs Locally Inference runs entirely in your browser using WebGPU or WASM. No server required. Data Stays Private Your text and audio never leave your device. Zero data collection, zero tracking. Open Source TTSLab is MIT licensed and fully open source. Contribute models, report bugs, or build on top of the project.