Skip to main content
Back

tiny-voca Squeezes Full Speech Synthesis Into an 18 MB Open-Source Model

A 4.5M-parameter English TTS model runs 70x real time on CPU, trained from scratch in under 8 hours on a single RTX 4090

  • tiny-voca is a 4.5M-parameter, 18 MB English TTS model running 70x real time on CPU
  • Two American English voices (F2 female, M2 male) at 24 kHz, runs in-browser via ONNX Runtime Web
  • Uses a source-filter synthesizer with only 25,862 parameters instead of a neural vocoder
  • Trained from scratch in 7.7 hours on a single RTX 4090 across 104,737 steps
  • Training audio is fully synthetic, generated with Qwen3-TTS 1.7B Base and screened by Whisper
  • Community extended the recipe to 20 additional languages within one week; Apache-2.0 code and weights

Open-source project tiny-voca packages an English text-to-speech model, waveform generator included, into 4.5 million parameters. The repository passed 800 GitHub stars during its first week, and contributors have trained separate community checkpoints for roughly 20 additional languages.

The compact model targets local applications that cannot accommodate a large neural vocoder or cloud speech API. It runs through PyTorch, ONNX Runtime, or ONNX Runtime Web, enabling deployment on CPUs and inside a browser tab.

An 18 MB speech stack

  • Size: 4,537,291 inference parameters and an 18 MB checkpoint, including waveform synthesis.
  • Reported speed: about 70 times faster than real time with four CPU threads. At that rate, ten seconds of audio takes roughly 0.14 seconds to generate, although hardware and runtime choices affect performance.
  • Voices: F2 and M2, described by the project as female and male American-English voices.
  • Audio: 24 kHz output.
  • Runtimes: PyTorch, ONNX Runtime, and ONNX Runtime Web.
  • Licensing: code, weights, and training recipe under Apache 2.0; the released dataset under CC BY 4.0.

tiny-voca’s model weights were initialized from scratch, without a larger teacher model or pretrained neural vocoder. Its training audio was generated by Qwen3-TTS, so the from-scratch claim applies to model training rather than data production.

Classical DSP cuts the parameter count

tiny-voca follows a FastSpeech 2-style, non-autoregressive pipeline. A text encoder processes phonemes, after which two frame-level Transformers predict duration, pauses, pitch, voicing, and energy. The Transformer blocks use pre-normalization and rotary position embeddings.

A source-filter synthesizer replaces the neural vocoder. The model generates a band-limited pulse train at the predicted pitch alongside a noise signal, then shapes both through a time-varying filter. The network predicts the filter’s cepstrum, a compact representation of its frequency response, once per frame.