tiny-voca is an English text-to-speech system designed to run efficiently on ordinary CPUs and in browsers. Its complete model has about 4.54 million parameters, produces 24 kHz audio, and supports two fixed American English voices, one female and one male. On four CPU threads it runs roughly 70 times faster than real time. The model converts phonemes into speech using a transformer-based network that predicts pronunciation timing and prosody, then drives a compact source-filter waveform synthesizer rather than a neural vocoder. Its weights and code are released under Apache-2.0, with PyTorch and ONNX versions available.
The model was trained from scratch for 104,737 steps on 17.6 hours of synthetic speech, generated using Qwen3-TTS; no real-speaker recordings or pretrained model weights were used. Training, including experiments and failed runs, cost $9.56 on a rented RTX 4090. In a November 2026 leaderboard snapshot, its male voice ranked eighth and female voice thirteenth among 50 English systems by average word error rate, though the developer’s own tests are small and no formal listening study was conducted. Speech is generally clear but can sound synthetic; the system is English-only, offers no voice cloning or streaming, and may mispronounce unusual names and spellings. Prosody is fixed, and the female voice sounds younger than intended.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.