Skip to content
Audio

TinyVoca DDSP

ultra-lightweight CPU speech synthesis without PyTorch

2 min read

TinyVoca DDSP by TuringLogic is a compact 4.65M-parameter English text-to-speech (TTS) model that outputs 24 kHz mono audio. Built on the TinyVoca architecture, it replaces the standard vocoder with DDSP (Differentiable Digital Signal Processing using 64 harmonics and filtered noise) and was trained from scratch on 29.6 hours of real speaker recordings from the HiFi-TTS dataset. Unlike most modern TTS systems, it requires neither a GPU nor bulky PyTorch: the model weighs around 25 MB including pronunciation dictionaries and runs directly on a CPU or in the browser via ONNX Runtime. It's built for developers creating local voice interfaces, embedded systems, and web apps where a tiny footprint and instant CPU response are critical.

Key features

  • Tiny footprint: just 4,648,907 parameters; the FP32 ONNX model takes up 19.52 MB (around 25 MB with pronunciation dictionaries).
  • Fast CPU inference: on an Apple M4 Pro, generation via ONNX Runtime 1.32.0 runs at 109.9x real-time speed (RTF 0.0091), synthesizing the first phrase in 0.115 seconds.
  • No PyTorch dependency: inference only requires NumPy, ONNX Runtime, SoundFile, and the dizione phonemizer, with browser execution also supported via ONNX Runtime Web.
  • Two voices and tempo control: includes two speaker voices (92 and 6097), speech rate adjustment, and CPU thread count control.

What you can do

  • Embed voice synthesis directly into a web page: run the model client-side via ONNX Runtime Web with zero server GPU costs (web demo source code is included in the repo).
  • Add offline speech to a local script or bot: voice short English notifications and replies on a low-end server or laptop with around 500 MB peak RAM usage.
  • Train on your own voice: use the open-source PyTorch training pipeline in the repository to train your own compact DDSP model.

Example

hf download TuringLogic/tiny-voca-ddsp --local-dir model \
  --include 'onnx/*' 'phonemizer/*' 'onnx_infer.py' 'requirements-onnx.txt'
pip install -r model/requirements-onnx.txt
python model/onnx_infer.py --voice 92 \
  --text "Hello, this runs on a CPU." --output hello.wav
Read moreDetails
Intelligibility (mean sentence WER via faster-whisper small.en): 13.45% on validation (vs. 4.85% on ground-truth recordings), 7.33% on test, and 1.67% on unseen phrases.
Limitations: English only, single request capped at 512 phoneme tokens and 3,000 acoustic frames (around 32 seconds of audio), full-utterance generation without streaming.
Training: 104,000 steps (37,833 audio clips, ~3h 27m on the training server). Weights are released under the non-commercial CC BY-NC 4.0 license; code is under Apache-2.0.
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # TuringLogic/tiny-voca-ddsp
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download TuringLogic/tiny-voca-ddsp --local-dir ~/models/tiny-voca-ddsp
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to tiny-voca-ddsp
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of tiny-voca-ddsp (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

TuringLogic/tiny-voca-ddsp
Author
TuringLogic
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Audio

Sayso

Sayso
Audio

fiato

fiato