Developer Muhammed Nazeem, who posts as n4ze3m, introduced tiny-voca in a November 15 post. The English text-to-speech model has 4.5 million parameters, roughly 18 MB of weights and support for CPU and browser use, according to its documentation. The GPU can sit this reading out. SourceSourceSource
“I trained a text-to-speech model that has never heard a real human voice,” Nazeem wrote. Tiny-voca’s own training audio was synthetic; the project’s documentation doesn’t establish whether its upstream models trained on human recordings. SourceSource
According to the dataset documentation, its two voices were created from text descriptions using Qwen3-TTS VoiceDesign, rather than recordings or imitations of real people. Qwen3-TTS 1.7B Base then generated the speech used to train tiny-voca. SourceSource
Small weights, specific limits
The model card lists exactly 4,537,291 parameters, including the waveform generator. Its PyTorch weights occupy 18.2 MB, while the ONNX export used for browser operation is 21.4 MB. The text-to-phoneme system also has 5.4 MB of data files. So the roughly 18 MB figure describes the PyTorch weights, not an all-inclusive browser download. Source
Nazeem reports generation at about 70 times real time using four CPU threads. The documentation doesn’t identify the CPU, provide a reproducible timing protocol or include independent verification. That isn’t a speed guarantee for whichever laptop happens to be nearby. SourceSource
The model offers two fixed American English voices, F2 and M2, with 24 kHz mono output and no voice cloning. The model card warns that speech can sound slightly synthetic, foreign names and unusual spellings are often mispronounced, and identical text gets the same delivery each time. Two voices, not an audition department. Source
The project’s small evaluation uses speech recognition to check generated speech. The model card calls the scores a sanity check, not a benchmark, and says no formal listening test has been run. Those scores don’t establish how natural the voices sound. Source
Open code, synthetic lessons
For waveform generation, tiny-voca uses a source-filter synthesizer instead of a neural vocoder. In the model card’s description, the network predicts filters that shape a pulse signal and noise into audio. Source
The released model was trained on 17.6 hours of synthetic speech, with the final training run taking 7.7 hours on one RTX 4090, according to the model card. It contains no pretrained weights, although pretrained models were used to generate and label its training data. Source
The project provides code and a training recipe, weights and a dataset. Code and weights are released under Apache-2.0, while the dataset is CC BY 4.0; phonemizer data carries separate licenses. SourceSource
The repository also links to a browser demo for trying the two voices locally. Source
Comments