18MB browser-run TTS model tiny-voca makes paid English TTS look silly
thisdudelikesAI · x · 2026-11-17
tiny-voca, a 4.5M-parameter English TTS model, runs entirely in your browser from an 18MB file (21MB ONNX).
- 4,537,291 params including the waveform generator, trained from scratch on 17.6 hours of synthetic speech
- 70x real time on a CPU via PyTorch/ONNX Runtime, 24kHz mono, two fixed American English voices
- Nothing you type leaves your device; Apache 2.0 code and weights
No GPU, account, or API needed. Limits: English-only, two fixed voices, no cloning, no streaming, and foreign names often misread — but the author argues it makes paid English TTS look silly.
Related event: tiny-voca: 18MB TTS model runs in browser at 70x real-time on CPU(4 posts)→
More from Infra
- A 10-item beginner reading list for LLM distributed training, from Horovod to FSDP — East-Muffin-6472 · 2026-11-17
- Replicate quietly switches nano banana outputs from JPG downloads to presigned URLs — Kyrannio · 2026-11-17
- Fireworks AI tells investors chips are a commodity, its moat is software that runs models fast — hargup13 · 2026-11-17
- Claude-written CUDA C++ port runs TRELLIS.2 5.3x faster on H200, bit-identical — huggingface · 2026-11-17
- China to make 3M AI ASICs in 2027 vs West's 20M, each ~4x faster per chip — teortaxesTex · 2026-11-17
- Harmonic's AI theorem prover Aristotle takes a no-hallucination hardware synthesis talk to Cambridge — satnam6502 · 2026-11-17