Tiny Voca German/English

Two independent fixed-roster models trained from random initialization using the tiny-voca architecture and locally generated fictional Qwen voice supervision. This is not a fine-tune of upstream tiny-voca pretrained weights.

Variant Parameters Safetensors bytes Voices
voca8/ 4,541,899 18,186,220 4 German, 4 English
voca64/ 4,563,403 18,272,244 32 German, 32 English

Both finished a fixed 104,000-step run. Equal steps do not mean equal exposure per voice or an isolated capacity comparison. Voice embeddings are checkpoint-local, even when their labels match.

Reference inference

Source and tests: https://github.com/fraunhofer-iem/designed-voices-tts Initial verified public source revision: e1158b4e47053f2dae656e9c7ba1535554c7780c.

Install the reference package from that pinned public archive:

uv venv
uv pip install 'designed-voices-tts @ https://github.com/fraunhofer-iem/designed-voices-tts/archive/e1158b4e47053f2dae656e9c7ba1535554c7780c.tar.gz' huggingface-hub

Download only the desired variant, then use its safe loader:

from huggingface_hub import snapshot_download
from designed_tts.release import load_release
root = snapshot_download("Fraunhofer-IEM/tiny-voca-de-en",
                         allow_patterns=["voca64/*"])
model = load_release(root + "/voca64")
print(model.voices)
# model.synthesize(phone_ids, voice, stress=stress_ids) -> CPU waveform, duration metadata

Inputs are the append-only 88-symbol phoneme/stress ABI, not raw text or reference audio. Complete text inference requires nrmlz and Phony; the public browser pipeline is still being qualified. Output is 24 kHz mono. Configs pin the original checkpoint and safe-weight hashes. Default neutral noise seed: 20261119.

Limits and verification

No zero-shot cloning, reference-audio inference or reliable emotion control. Quality varies by voice; lower-register output can sound metallic or phasey. Listen before production use. Do not treat a numerical/ASR score as naturalness.

All tensors survive export exactly. Four fixed native cases per variant matched the original waveform sample-for-sample. This is not ONNX/WASM, phone, streaming TTFA or minimum-RAM qualification. The sizes above exclude frontend/runtime assets.

Training data: https://huggingface.co/datasets/Fraunhofer-IEM/designed-voices-de-en The owner's public dataset upload is in progress; no completed public revision is claimed to reproduce the entire frozen training subset yet. Full training tooling and the client Space remain in progress.

See LICENSE, ATTRIBUTION.md and THIRD_PARTY_NOTICES.md for code, initialization, teacher, text and dictionary credits. No private rankings or human recordings are included.


Gefördert durch das Bundesministerium für Forschung, Technologie und Raumfahrt (BMFTR) im Rahmen des Projekts ENGAGE-KI. Betreut vom Projektträger Karlsruhe (PTKA). https://engage-ki.de/

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train Fraunhofer-IEM/tiny-voca-de-en

Space using Fraunhofer-IEM/tiny-voca-de-en 1