n4ze3m · Released, not on our provider

Tiny Voca (2026): what the 4.5M-parameter English text-to-speech model is, its voices, license, speed and how to run it

Tiny voca is a small open-source English text-to-speech model by Muhammed Nazeem (n4ze3m) with 4,537,291 parameters and two American English voices, F2 and M2, that writes 24 kHz audio, runs at about 70x real time on four CPU threads by its maker’s own figure, and is released under Apache-2.0 on Hugging Face.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Qwen3.8 Flash, not tiny voca.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
n4ze3m
Type
Text
Released
2026-11-14
OpenRouter id
None

Checked

tiny voca: facts, with sources

  • Hugging Face’s record of the n4ze3m/tiny-voca repository shows it was created on November 14, 2026 (UTC) and last modified on November 18, 2026; the GitHub repository n4ze3m/tiny-voca was created the same day. The model card gives no separate release date, and its citation block is dated 2026, so we say "repository created" rather than "released".Source: huggingface.co · checked 2026-11-19
  • The model card calls tiny-voca a small English text-to-speech model with two voices, F2 (female) and M2 (male), both American English. The whole model, including the waveform generator, has 4,537,291 parameters ("4.5M"), was trained from scratch on about 18 hours of synthetic speech and writes 24 kHz mono audio. Input is text, which is turned into phonemes outside the model by the author’s dizione tool.Source: huggingface.co · checked 2026-11-19
  • By the maker’s own figure (not tested by us), tiny-voca runs about 70x faster than real time on four CPU threads. The weights come as a 18.2 MB PyTorch file (float32) and a 21.4 MB ONNX export (float32, opset 17), plus 5.4 MB of phonemizer data; the same ONNX file runs in onnxruntime-web (WebAssembly), where the card says the audio matches the PyTorch model to within 0.6% of the peak amplitude, apart from the random noise signal.Source: huggingface.co · checked 2026-11-19
  • License: the model card says the weights and code are released under Apache-2.0. The files in the phonemizer folder are derived from the CMU Pronouncing Dictionary and other sources and keep their own licenses (LICENSE.cmudict and LICENSE.homograph-data in the repository). The GitHub README adds that the training dataset is CC BY 4.0. The card states no extra commercial-use limit.Source: huggingface.co · checked 2026-11-19
  • The card’s own limits list: English only, two fixed voices and no voice cloning; the voice is clear but can sound slightly synthetic; F2 sounds younger than the adult voice it was designed to be; the same text is always read the same way (no sampling of prosody); speech is slightly faster than the training recordings at the default speed; foreign names and unusual spellings are often mispronounced; there is no streaming; and because the training data is itself synthetic, the model cannot be more natural than the model that generated it.Source: huggingface.co · checked 2026-11-19
  • The repository file list on Hugging Face has the PyTorch weights (tiny-voca.pt), the ONNX file, config.json, the phonemizer data, eight sample WAV files and a gguf folder with five GGUF files (f16, q4_0, q8_0 and two others); the card does not describe the GGUF files. The GitHub README lists projects other people built on tiny-voca, such as ports to other languages and a run through llama.cpp from GGUF files, which we did not check.Source: huggingface.co · checked 2026-11-19
  • The author is named on the card as Muhammed Nazeem (Hugging Face and GitHub name n4ze3m), an individual, not a company. The card says the whole project cost $9.56 of GPU time on one rented RTX 4090, covering the dataset, the failed tries and the final run, and the GitHub README calls the code, weights, training recipe and dataset all open.Source: huggingface.co · checked 2026-11-19
  • tiny voca is not on OpenRouter’s model list (checked November 19, 2026), so this site cannot run it, and this site’s chat is text in, text out: it does not generate speech or audio. For text chat with small models use Qwen3.8 Flash or Inkling 2 Small instead.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

tiny voca specs: parameters, voices, audio, license and training

  • Architecture as the card describes it: text becomes phonemes with a stress value each; a transformer encoder of 4 blocks reads them, a duration predictor sets each phoneme’s length in frames, and two frame-level transformers (2 and 5 blocks) predict pitch, voicing and energy and then the waveform generator’s parameters. All blocks use 192 channels, 4 heads and rotary position embeddings. The waveform comes from a source-filter synthesizer with 25,862 parameters instead of a neural vocoder.Source: huggingface.co · checked 2026-11-19
  • Training as the card and the GitHub model notes describe it: 16,018 clips (17.6 hours) in two voices, 104,737 steps on one RTX 4090 in 7.7 hours. The audio is synthetic: both voices were designed from text descriptions with Qwen3-TTS VoiceDesign and the recordings were generated with Qwen3-TTS 1.7B Base, so no recording of a real speaker is involved. Phoneme timings come from forced alignment with facebook/wav2vec2-lv-60-espeak-cv-ft, used only to label the training data. The dataset page counts 17,371 clips and 19.1 hours across two batches.Source: github.com · checked 2026-11-19

Spec table

tiny voca at a glance, as printed on its model card

Most "tiny voca" searches are about size: how small can a speech model be and still sound usable. The card’s answer is 4.5M parameters and 18 MB of weights, with a source-filter waveform generator instead of a neural vocoder. It is a tool for reading text aloud on a CPU or in a browser, not a chat model, so it cannot replace a model like Qwen3.8 Flash. Another small model that runs on your own device, for text, is Aion 1.0. A Persian voice model with cloning is on the Cilla Mini page.

Our table: figures as printed on the tiny-voca Hugging Face card and the maker’s GitHub notes, read on November 19, 2026; they are the maker’s own figures and we did not run them · checked 2026-11-19
ItemAs printed
Parameters4,537,291 (4.5M), waveform generator included
VoicesF2 (female) and M2 (male), American English; no voice cloning
Output audio24 kHz mono
InputText, converted to phonemes with dizione (English only)
Weight filesPyTorch 18.2 MB; ONNX 21.4 MB (float32, opset 17); phonemizer data 5.4 MB
Speed (maker’s figure)About 70x faster than real time on four CPU threads
Training data16,018 synthetic clips, 17.6 hours, two voices
Training run104,737 steps, one RTX 4090, 7.7 hours; $9.56 of GPU time for the whole project
LicenseApache-2.0 for weights and code; phonemizer data keeps its own licenses; dataset CC BY 4.0

Source: huggingface.coSource: github.comSource: github.com

Our table

tiny voca quality: the maker’s word error rates and the Open TTS Leaderboard rows

  • The maker’s own check (not run by us): word error rate of synthesized speech transcribed with Whisper small.en was 0.6% on 8 validation clips and 0.9% on 8 test clips, the same as the recordings of those sentences, and 0.0% on 16 new sentences. The card says these sets are small, the scores are a sanity check and not a benchmark, and no formal listening test has been run.Source: huggingface.co · checked 2026-11-19
  • The card prints a snapshot of the Open TTS Leaderboard dated November 16, 2026 (v1.6.3): English, without voice cloning, equal average across Seed-TTS, CV3-Eval and MiniMax, 50 ranked models. tiny-voca (M2) ranks 8th by average word error rate (0.733 %, UTMOS22 4.34) and tiny-voca (F2) 13th (0.773 %, UTMOS22 4.43); each voice is a separate row and the leaderboard notes tiny-voca is CPU-only. We copied the figures from the card and did not open the leaderboard.Source: huggingface.co · checked 2026-11-19

Leaderboard excerpt

tiny voca next to the models ranked around it on the Open TTS Leaderboard (snapshot of November 16, 2026)

A model this small is easy to dismiss, so the card prints where it sits on a public leaderboard: the male voice M2 is 8th of 50 by average word error rate, the female voice F2 is 13th. The table shows the neighbours the card lists. Word error rate says how many words an automatic recognizer got wrong when it listened to the speech; UTMOS22 is an automatic score of naturalness. Neither replaces listening. The card itself says no formal listening test has been run. If you came here wanting a chat model, try Inkling 2 Small on this site; tiny voca only reads text aloud.

Our table: rows copied from the excerpt printed on the tiny-voca Hugging Face card (English, no voice cloning, average of three test sets, lower word error rate is better), read on November 19, 2026; we did not open the leaderboard and we did not run any model · checked 2026-11-19
ModelRankParamsAverage WER % (lower is better)UTMOS22 (higher is better)
phonon-31m-en641 M0.7304.45
Inflect Micro v3712 M0.7304.37
tiny-voca (M2)84.5 M0.7334.34
phonon-40m-en951 M0.7374.45
supertonic-310100 M0.7434.45
PiVo-1.4M121.4 M0.7674.33
tiny-voca (F2)134.5 M0.7734.43
LFM3-Audio-2B142.05 B0.7974.45

Source: huggingface.co

Our table

How to run tiny voca: PyTorch, ONNX, browser and plain JavaScript

  • Limits of one call, from the card: the ONNX graph reads one utterance per call, up to 512 tokens and 32 seconds, so longer text has to be split into sentences (the Python command does this by itself); end each utterance with a mark such as a full stop or the last word breaks off; and the audio can peak a little above 1.0 (about one sentence in twenty), so scale it down before saving 16-bit samples.Source: huggingface.co · checked 2026-11-19

Ways to run it

The ways the maker documents to run tiny voca and what each needs

tiny voca is meant to run on your own machine, so there is nothing to sign up for. The shortest route is the PyTorch command: install dizione and the tiny-voca package from GitHub, download the weights with the hf tool, set the DIZIONE_DATA variable to the phonemizer folder and call the inference module with a voice and a text. The ONNX route needs onnxruntime, soundfile and NumPy and gives you the same audio from a plain graph. Each row below says what you need. This site does not run it, because Qwen3.8 Flash and the other models here answer in text, not audio.

Our table: from the Hugging Face card and the GitHub README, read on November 19, 2026; we did not run any of these commands · checked 2026-11-19
RouteHowWhat it needs
PyTorch command linepython -m tiny_voca.infer with --checkpoint, --voice (F2 or M2), --text and --output; options for speed, pitch shift and pitch rangepip install of dizione and tiny-voca from GitHub, the weights from Hugging Face, DIZIONE_DATA set
ONNX Runtimeonnx/model.onnx takes tokens, stress, voice (0 = F2, 1 = M2), speed, pitch_shift and pitch_range and returns audio at 24 kHzonnxruntime, soundfile, NumPy and dizione for the phoneme ids; one utterance per call
BrowserThe maker’s demo at tiny-voca.n4ze3m.com runs the model in the browser and plays each sentence as soon as it is madeA modern browser; onnxruntime-web (WebAssembly) runs the same ONNX file
Plain JavaScriptAn experiment in the js folder: the model and phonemizer in plain scripts, one 8 MB file, in the browser, Node and BunNode or Bun, or a browser; no libraries

Source: huggingface.coSource: github.com

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to tiny voca you can use now

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

Inkling 2 Small

Thinking Machines Lab

$0.30 / $1 (per million tokens)

  • Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
  • A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
  • Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14

FAQ

tiny voca FAQ

What is tiny voca?

Tiny voca is a text-to-speech model: you give it English text and it returns a spoken audio file. Its card calls it a small English model with two voices, F2 (female) and M2 (male), 4,537,291 parameters and 24 kHz mono output, trained from scratch on about 18 hours of synthetic speech. It is not a chat model and it does not write text. See the table for the full spec.

Who made tiny voca and when did it appear?

The card names Muhammed Nazeem, who posts as n4ze3m, as the author; it is an individual project, not a company product. Hugging Face shows the repository was created on November 14, 2026 and last modified on November 18, 2026, and the GitHub repository was created the same day. The card gives no separate release date.

What is the license, and can I use tiny voca in a product?

The card says the weights and code are released under Apache-2.0, and the GitHub README says the training dataset is CC BY 4.0. The phonemizer data comes from the CMU Pronouncing Dictionary and other sources, each with its own license file in the repository. The card states no extra commercial limit, but read the license files yourself before you ship; this page is not legal advice.

How fast is tiny voca and what hardware does it need?

By the maker’s own figure, which we did not test, it runs about 70x faster than real time on four CPU threads, so it needs no GPU. The weights are 18.2 MB (PyTorch) or 21.4 MB (ONNX), and the card says the same ONNX file runs in a browser through onnxruntime-web. Each call handles one utterance of up to 512 tokens and 32 seconds, so long text is split into sentences.

How good does tiny voca sound?

The card says the voice is clear but can sound slightly synthetic, and that no formal listening test has been run. Its word error rate from Whisper small.en is 0.6% on validation and 0.9% on test clips, and the card prints leaderboard ranks of 8th (M2) and 13th (F2) among 50 English models without voice cloning. Those are the maker’s figures; listen to the demo before you decide.

Can tiny voca clone a voice or speak other languages?

No. The card’s limits say English only, two fixed voices and no voice cloning. Both voices were designed from text descriptions and the training audio is synthetic, so no real speaker is involved. The GitHub README lists community projects for other languages built on tiny voca; we did not check them. For a voice model that does clone, see Cilla Mini, which speaks Persian.

Can I use tiny voca on this site?

No. It is not on OpenRouter’s model list and this site does not generate speech; its chat answers in text. You can chat with a small model instead: on this site a typical reply to Inkling 2 Small costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Qwen3.8 Flash is another.

Is this the official tiny voca website?

No. This page is unofficial and not affiliated with or endorsed by n4ze3m. Facts link to the pages we read.