Wink Pings

A 4.5M-parameter speech model trained for $9.56: why are people comparing it to Apple transcription?

tiny-voca uses 18MB to run 70x real-time speech synthesis on a CPU, and training it overnight cost $9.56. Gregor Zunic said it is better than Apple transcription, and the comments promptly corrected the difference between TTS and STT.

tiny-voca, a 4.5M-parameter English TTS model, was trained for 7.7 hours on a single RTX 4090, with a total project cost of $9.56. It is only 18MB, can run on a CPU, and is about 70x real-time.

It was made by Nazeem. The code, weights, dataset, and training recipe have all been released under Apache-2.0. The demo page runs directly in the browser, and the text you enter never leaves your device. The model comes with its own phonemizer and can read numbers, dates, and currencies. Nazeem gives examples: “RTX 4090” is read as “forty ninety,” and “$4,090” is read as “four thousand and ninety dollars.”

When Gregor Zunic reposted it, he said that a 4.5M model trained overnight is much better than Apple transcription, and asked what Apple is doing. The comments quickly corrected him: this is TTS, while Apple transcription is STT. They are two different tasks. Some also pointed out that making ASR robust across a wide range of accents requires more data and capacity, and you cannot directly compare it with a TTS model fixed to two voices.

Another point of contention was “distillation.” Some comments said it was not trained from scratch. The model card states it clearly: there are no pretrained weights, and the training data is 17.6 hours of synthetic speech, 16,018 clips, generated by Qwen3-TTS. If that is called distillation, then what is being distilled is the data, not the weights.

The limitations are also clearly listed: English only, two fixed voices F2 and M2, no voice cloning, no streaming, the same text is always read the same way, foreign names and uncommon spellings are often mispronounced, and the F2 voice sounds younger than intended. The training data itself is synthetic, so it cannot be more natural than the model that generated the data.

But 18MB running on a CPU means an app can embed offline speech synthesis directly, without API calls. The size itself is the message. The fact that Apple got pulled into the comparison shows that the timeline's expectations for small local models have already changed. When the comparison target is wrong, the discussion goes off track.

If you want to try it, open the browser demo and type in text: https://tiny-voca.n4ze3m.com The model is on HuggingFace: https://huggingface.co/n4ze3m/tiny-voca The code is on GitHub: https://github.com/n4ze3m/tiny-voca

tiny-voca成本 breakdown

发布时间: 2026-11-16 20:32