Spaces:
Runtime error
title: tiny-voca
emoji: 🗣️
colorFrom: pink
colorTo: purple
sdk: gradio
sdk_version: 6.32.0
app_file: app.py
python_version: '3.12'
short_description: 4.5M-param English TTS, two voices, 24 kHz, runs on CPU
tiny-voca
A demo of n4ze3m/tiny-voca: a 4.5M-parameter English text-to-speech model — two voices (F2 female, M2 male), 24 kHz mono audio, ~70× real-time on CPU. Weights and phonemizer data are Apache-2.0.
Text is turned into phonemes by the author's own dizione phonemizer (vendored in this Space): a CMUdict-derived lexicon, a small neural G2P model for unknown words, and a homograph classifier — so "I read the book you read yesterday" comes out right.
The ONNX export of the model (21 MB) is downloaded from the model repo at startup and run on CPU via onnxruntime. Long text is split into sentences, long sentences into utterances that fit the model's 512-token limit, and the resulting audio chunks are concatenated — the same pipeline as the author's browser demo.