← Back to all posts
Tools

tiny-voca, a $9.56 Open TTS Model, Ranks 8th of 50 Ahead of a 2B-Parameter Rival

November 17, 2026 · 00:12 UTC · Tools
On this page 8 sections
tiny-voca, a $9.56 Open TTS Model, Ranks 8th of 50 Ahead of a 2B-Parameter Rival

TL;DR

tiny-voca is a 4.5M-parameter English text-to-speech model that its author, Muhammed Nazeem, trained from scratch for $9.56 of GPU time on one rented RTX 4090. Code, weights and training data are all public, Apache-2.0 for the first two. The training set contains no recording of a real speaker: 17.6 hours of audio in two voices, generated end to end by a larger TTS model. On the Open TTS Leaderboard snapshot dated 16 November 2026, the male voice ranks 8th of 50 by average word error rate, ahead of entries up to roughly 450 times its size, and inference runs about 70x real time on four CPU threads.


At a glance
A full TTS model for the price of lunch
4.5M
parameters, waveform generator included
$9.56
total gpu spend, dataset and failed runs included
8th
of 50 by average wer, male voice
70x
real time on four cpu threads
18.2MB
pytorch weights, 21.4MB as onnx
0
recordings of real speakers in the training set

What shipped

The whole artifact is small enough to read in an evening. The GitHub repo holds training code, tests and a web demo; the weights are an 18.2 MB PyTorch file plus a 21.4 MB ONNX export. The dataset ships under CC BY 4.0, and the phonemizer, dizione, is its own small project you can reuse. There is also a browser demo that runs the ONNX graph in WebAssembly; per its page, nothing you type leaves the device. At the time of writing the repo sat at 609 stars and 72 forks, with 185 likes on the model page.

Running it locally is four lines from the README:

git clone https://github.com/n4ze3m/tiny-voca.git && cd tiny-voca
pip install -e ".[own]"
hf download n4ze3m/tiny-voca --local-dir weights
python -m tiny_voca.infer --checkpoint weights/tiny-voca.pt --voice F2 --text "Hello! This is tiny voca." --output hello.wav

8th of 50, with an asterisk the card wrote itself

Each voice ranks as its own row on the leaderboard. In the 16 November snapshot, M2, the male voice, sits 8th of 50 English models with 0.733 average WER; F2, the female voice, sits 13th at 0.773. The only entries ahead of M2 are larger: Inflect Micro v3 at 12M and phonon-31m-en at 41M. Below it sit phonon-40m-en (51M), supertonic-3 (100M) and LFM3-Audio-2B at 2.05B parameters, about 450x tiny-voca's size. One smaller model, PiVo at 1.4M, lands between the two voices, so the smallest model on the board did not come from this project.

Open TTS Leaderboard, 16 Nov snapshot
Average WER: a tight field, lower is better
rivalstiny-voca
inflect 12m
0.730
tiny m2
0.733
phonon 40m
0.737
supertonic
0.743
pivo 1.4m
0.767
tiny f2
0.773
lfm3 2b
0.797
Average WER across the board's three English eval sets; the entire spread is under 0.07 points. Source: Open TTS Leaderboard snapshot in the tiny-voca model card, 16 Nov 2026.

Read the metric before you quote the rank. WER here means a Whisper model transcribes the synthesized speech and someone counts the wrong words, so it measures intelligibility, with naturalness only approximated by a second automatic score (UTMOS22 puts the two voices at 4.34 and 4.43 against a top of about 4.45). The card cops to its own limits in plain text: the WER sanity checks use 8, 8 and 16 sentences, and "the scores are a sanity check, not a benchmark. No formal listening test has been run." The hosted leaderboard also notes both tiny-voca rows are CPU-only.

How 4.5M parameters make a voice

The design descends from FastSpeech 2, shrunk hard. Text goes through dizione to phonemes with stress marks. A 4-block transformer encoder (192 channels, 4 heads, rotary embeddings) reads them, a duration predictor assigns each phoneme a length in frames, and two frame-level transformers predict pitch, voicing and energy. All standard. The unusual part is the final 25,862 parameters.

Most modern TTS ends in a neural vocoder, a network that paints the waveform sample by sample. tiny-voca ends in something closer to a synth patch: a band-limited pulse train at the predicted pitch and a noise signal, each pushed through a time-varying filter whose shape the network predicts once per 256-sample frame, every 10.7 ms. A vocoder is a tape recorder; this is an instrument the model retunes about 94 times a second. That swap is a large share of why the whole thing fits in 18 MB and runs 70x real time on CPU.

From text to waveform
Five stages, no autoregressive loop
Input
Text to phonemes
dizione adds stress marks, sentences split
Encode
4-block transformer
192 channels, 4 heads, rotary embeddings
Stretch
Duration predictor
each phoneme gets a length in 256-sample frames
Prosody
Two frame transformers
pitch, voicing and energy, predicted in parallel
Output
Source-filter head
25,862 params emit 24 kHz mono audio
Training used multi-resolution spectral losses, with adversarial losses joining after step 8,000. One frame is 10.7 ms.

The training data never met a human

Both voices were designed from text descriptions with Qwen3-TTS VoiceDesign, and all 16,018 clips of training audio were then generated by Qwen3-TTS 1.7B Base. In the card's words, "No recording of a real speaker is involved." Forced alignment with a wav2vec2 model labeled the phoneme durations, and it was used for nothing else. Training ran 104,737 steps in 7.7 hours.

That data choice does two jobs at once. It makes the project cheap to reproduce, since the dataset is public and the teacher is a commodity API away, and it removes the entire question of performer consent and licensing, because there are no performers. The ceiling is equally clear, and the card states it: the model "cannot be more natural than the model that generated it."

The whole bill
What $9.56 bought
17.6h
synthetic training audio, two voices
16,018
clips in the dataset
104,737
training steps
7.7h
on one rented RTX 4090
$9.56
dataset, failed tries and the final run
The author's own accounting, from the model card's cost section. His summary: "It cost less than a pizza to make."

The internet ported it within days

The repo's Community section, which the author says he has not tested, already lists ports to German (apptor-tiny-tts), Italian (tiny-voca-it), Turkish (zerre-tts) and Korean (korean-tts), a rewrite in Rust, a Mac app, a local dashboard, and tinygroove, a drum machine with no samples built on the same idea. fiato, an English variant from another developer, adds sighs, exhales and gasps.

Should you ship with it

For anything that needs a voice where there is no GPU and no network call: embedded toys, browser apps, homelab readouts, prototypes of talking products. The fixed inventory is two American English voices, no cloning, which is also why the licensing story is so clean. The card's Limitations section is longer than most marketing pages, and usefully so: foreign names and odd spellings often come out wrong, numbers and abbreviations depend on the phonemizer, there is no streaming, the same text is always read the same way, and F2 was designed as a woman in her mid-twenties and sounds younger, with a more mature voice planned. If you need a distinctive brand voice or multilingual coverage out of the box, this is the wrong shelf. If you need speech that runs anywhere for free, this one is worth the evening it takes to read.

Key Takeaways

  • tiny-voca is 4,537,291 parameters trained from scratch, no pretrained weights anywhere, with an 18.2 MB PyTorch file and a 21.4 MB ONNX export under Apache-2.0 and the dataset under CC BY 4.0.
  • Total project cost: $9.56 of GPU time on one rented RTX 4090, covering dataset generation, failed attempts and the 104,737-step final run of 7.7 hours.
  • All 17.6 hours of training audio were generated by Qwen3-TTS; no real speaker was recorded, so performer licensing never becomes a question.
  • Open TTS Leaderboard, 16 Nov 2026 snapshot: the male voice ranks 8th of 50 at 0.733 average WER and the female voice 13th at 0.773, ahead of a 2.05B-parameter model; the smallest entry ahead of it is 12M.
  • WER measures intelligibility via Whisper transcription; the card itself flags the eval sets as sanity checks with no formal listening test, so treat quality claims accordingly.
  • Inference runs about 70x real time on four CPU threads, and the ONNX graph runs in onnxruntime-web, proven by a browser demo that never sends your text anywhere.

Sources: tiny-voca model card, tiny-voca GitHub repository, Open TTS Leaderboard, tiny-voca dataset, browser demo

AITTStiny-vocaOpen SourceSynthetic DataSpeech SynthesisApache-2.0Local Inference
CONSOLE
$