tiny-voca-ru

A Russian adaptation of the 4.5M-parameter tiny-voca text-to-speech model with female and male voices, trained from scratch on egnch/tiny-voca-ru-dataset.

Property Value
Parameters 4,534,411 (~18.2 MB safetensors, ~21.4 MB ONNX)
Voices F (female), M (male)
Output 24 kHz mono
Frontend Pure-Python Russian normalizer, stress (lite / full), 61-symbol IPA
Weights model.safetensors + config.json, onnx/model.onnx + onnx/model.json, onnx/omograph.onnx
CPU Speed (4 threads) 115× real-time (PyTorch), 81× real-time (ONNX Runtime)
License Apache-2.0 (bundled ruaccent stress data: MIT)

Audio Samples

Voice Text Audio
F Здравствуйте! Это tiny-voca-ru, маленькая модель синтеза речи на русском языке.
M Здравствуйте! Это tiny-voca-ru, маленькая модель синтеза речи на русском языке.
F Встречу перенесли на 15:30, 14 марта, так что у нас осталось около 45 минут. Билет стоит 1250 рублей.
M Ты правда думал, что я забуду про твой день рождения? Конечно, нет!

Usage

Installation

# ONNX Runtime inference (no PyTorch)
pip install "tiny-voca-ru[onnx]"

# PyTorch + safetensors
pip install "tiny-voca-ru[torch]"

Python API

from tiny_voca_ru import Synthesizer

synth = Synthesizer.from_pretrained("egnch/tiny-voca-ru")
# ONNX Runtime backend
# synth = Synthesizer.from_pretrained("egnch/tiny-voca-ru", backend="onnx")

result = synth.save_wav(
    "Здравствуйте! Встречу перенесли на 15:30, 14 марта.",
    "hello.wav",
    voice="M",  # "F" or "M"
    speed=1.0,  # 0.25 .. 4.0
    pitch_shift=0.0,  # log-F0 offset (0.058 ≈ +1 semitone)
    pitch_range=1.0,  # 0.0 .. 3.0
    stress="lite",  # "lite" | "full" | "none"
)

CLI

tiny-voca-ru-speak   --voice F   --text "Привет! Билет стоит 1250 рублей."   --output out.wav

Pass --backend onnx to use ONNX Runtime. Manual stress can be marked with + before the stressed vowel (зам+ок).

Changes from tiny-voca

Follows the tiny-voca architecture and training recipe, adapted for Russian:

  • Phoneme inventory: 61-symbol Russian IPA set (5 special, 12 punctuation, 36 consonants with palatalization pairs, 8 vowels including reduced ɐ and ə).
  • Frontend (tiny_voca_ru.frontend): Pure-Python Russian text normalization, stress placement (lite via bundled stress.txt.gz + accent.onnx, or full via 10 MB INT8 onnx/omograph.onnx), and rule-based G2P.
  • Training: Trained from scratch for 104,000 steps (4.9 h on 1× NVIDIA L40S) on 18,517 aligned clips (21.96 h) from egnch/tiny-voca-ru-dataset, using facebook/wav2vec2-lv-60-espeak-cv-ft CTC durations during data preparation.

Custom-voice fine-tuning scripts (tiny-voca-ru-prepare, tiny-voca-ru-train) are available in the GitHub repository.

Evaluation

Evaluated at 104k steps with the default lite frontend against the Qwen3-TTS teacher on the test split (696 clips, Whisper large-v3-turbo ASR). CPU speed measured on AMD Ryzen 5 3600X (4 threads, 55.03 s of audio).

Metric Split tiny-voca-ru Teacher
CER test (696 clips) 0.46% (F 0.43%, M 0.49%) 0.31%
WER test (696 clips) 2.03% 1.54%
Stress accuracy (CTC top-1) test dict words (3,905) 98.5% 95.6%
Homographs (lite / ruaccent turbo3.1 ref) homographs (250) 56.0% / 91.6% —
Numbers ASR exact match test numeric templates (139) 90.6% 92.1%
Speed (PyTorch / ONNX CPU, 4 threads) 55.03 s audio 115× / 81× RT —
Speed (ONNX Runtime Web, WASM 1 thread) Chromium 16.7–21.7× RT —

Limitations

  • Russian only; two voices (F, M) inheriting Qwen3-TTS pronunciation traits.
  • Default lite stress does not disambiguate context-dependent homographs; use stress="full" (10 MB INT8 onnx/omograph.onnx) or explicit + marks.
  • Rule-based numeral case inflection and Latin transliteration can occasionally err on complex syntax or foreign names.

Citation

@misc{chetvertkov2026tinyvocaru,
  author = {Chetvertkov, Evgeniy},
  title  = {tiny-voca-ru: A 4.5M-Parameter Russian Text-to-Speech Model},
  year   = {2026},
  url    = {https://huggingface.co/egnch/tiny-voca-ru},
  note   = {ORCID: https://orcid.org/0009-0006-6158-2624}
}

If you use this model, please also cite the original tiny-voca:

@software{nazeem2026tinyvoca,
  author = {Nazeem, Muhammed},
  title  = {tiny-voca: a 4.5M-parameter two-voice English text-to-speech model},
  year   = {2026},
  url    = {https://huggingface.co/n4ze3m/tiny-voca}
}

License & Attribution

Downloads last month
50
Safetensors
Model size
4.53M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train egnch/tiny-voca-ru

Space using egnch/tiny-voca-ru 1