Field noteNov 17, 20265 min read1 source

tiny-voca is a 4.5M parameter English TTS model, 18 MB, running 70x faster than real time on CPU.

In one minute

Muhammed Nazeem released tiny-voca, an Apache-2.0 English text-to-speech model with 4.5 million parameters and two fixed voices. The project reports CPU generation at roughly 70 times real time, making it a small local option for constrained applications, while its narrow language and voice support and limited evaluation set keep the claim in perspective.

What changed: A separate English TTS release, tiny-voca, is much smaller at 4.5 million parameters and 18 MB, with two voices and reported CPU speed of 70× real time.

What happened

The tiny-voca model card describes a complete English TTS system with 4,537,291 parameters, F2 and M2 American English voices, and 24 kHz mono output. The PyTorch weights occupy 18.2 MB; an ONNX export is 21.4 MB. The license is Apache-2.0. The author says it runs about 70 times faster than real time on four CPU threads. It can also run in onnxruntime-web using WebAssembly, and the model card links to a browser demo. It is not deployed by an Inference Provider.

Text is converted to phoneme IDs and stress values outside the model using dizione. A four-block transformer encoder processes the phonemes; a duration predictor assigns lengths, followed by two frame-level transformers that predict pitch, voicing, energy and waveform-generator parameters. Instead of a neural vocoder, the system uses a source-filter synthesizer with a predicted time-varying filter. The ONNX graph accepts phoneme IDs, stress, voice, speed, pitch shift and pitch range, and returns a waveform plus token durations. The default speed is 1.0; the author notes that 0.058 pitch shift is about one semitone.

The system handles one utterance per call, up to 512 tokens and 32 seconds. The Python helper splits longer text sentence by sentence. The model card warns that a sentence should end with punctuation because an unpunctuated final word can break off. Saved audio should be scaled before writing 16-bit samples: output can peak above 1.0, which the author says happens about once in twenty sentences. The project is English-only, has no streaming or voice cloning, and always reads the same text the same way; it does not sample prosody.

Nazeem says tiny-voca was trained from scratch on 16,018 clips totaling 17.6 hours, with no pretrained weights. The two synthetic voices were created from text descriptions with Qwen3-TTS VoiceDesign, and their recordings were generated with Qwen3-TTS 1.7B Base. Training took 104,737 steps and 7.7 hours on one RTX 4090; the author puts total GPU rental cost, including failed attempts, at $9.56. For evaluation, Whisper small.en transcribed eight validation clips, eight test clips and 16 new sentences. The model card reports 0.6% and 0.9% word error rates on validation and test, respectively, and 0.0% on the new sentences, but explicitly calls these small sets a sanity check rather than a benchmark.

What the archive adds

New Runtime's earlier coverage includes Silencio-2, a compact English TTS model with emotion tags on 2026-11-01, Woof, a browser-focused TTS model on 2026-11-05, and ZONOS3-Mini, an on-device speech model on 2026-11-10. tiny-voca is a new release, not a new fact about those projects; it extends the compact open TTS thread with a smaller English model and CPU inference claim.

Trend signal

This release supports the broader shift toward generative media as application infrastructure: a small, licensed model with a browser-compatible runtime is easier to embed than a hosted, one-off generation service. It does not establish that compact speech systems can replace general-purpose voice stacks. New Runtime tracks this broader question in Generative media becomes application infrastructure.

Why it matters

For teams building offline-capable applications, demos, or simple spoken interfaces, tiny-voca offers a path to generate English speech without a remote inference dependency or a large model package. The source-filter output design and ONNX Web support are relevant because they make local execution practical across CPU and browser environments, while the small parameter count is not itself proof of production-ready quality.

The trade-off is a deliberately bounded product surface: two synthetic voices, English only, no streaming, no voice cloning, and limited prosody variation. The author also flags mispronunciation risks for foreign names and unusual spellings, and notes that the F2 voice sounds younger than intended. Teams should treat the release as a candidate for controlled, low-risk uses rather than a drop-in conversational voice layer. Test representative text, inspect audio peaks, and decide whether pronunciation and naturalness meet the application's bar before shipping.

Limits and uncertainty

The reported speed is the author's result on four CPU threads; the model card does not establish performance across other processors or workloads. Evaluation uses only 32 short examples, and the author says no formal listening test has been run. The voices are synthetic and derive from another TTS model, so their naturalness may inherit its limitations. There is no reported evaluation of pronunciation across varied names, domains, accents or long-form material. The model card describes planned work on a more mature-sounding F2 voice, but does not give a release date.

Reader takeaway

tiny-voca is notable as a very small, open English TTS system that can run locally on CPU or in a browser, not as evidence that small speech models have solved naturalness or broad coverage. Its best fit is a workload that values compactness and offline execution over streaming, voice variety or reliable pronunciation of unusual text. Validate those constraints with your own utterances before integrating it.

Evidence / sources

  1. 1huggingface.co/n4ze3m/tiny-voca
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicOpen ModelsExplore the open-models topic hub.
  2. 02topicGenerative MediaExplore the generative-media topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract