Skip to main content

tiny-voca-ru

English | Русский

A 4.5M-parameter Russian text-to-speech adaptation of tiny-voca (24 kHz mono, two synthetic voices F and M) with a pure-Python Russian text frontend, safetensors/ONNX Runtime inference, custom-voice fine-tuning, and a client-side browser demo.

Installation

Requires Python 3.10+. Works with pip or uv.

# Minimal Russian text frontend only (no PyTorch)
pip install tiny-voca-ru

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

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

# Full training & custom-voice fine-tuning stack
pip install "tiny-voca-ru[train]"

# Optional full ruaccent homograph disambiguation
pip install "tiny-voca-ru[ruaccent]"

Quick Start: Inference

Synthesizer loads models from Hugging Face Hub (egnch/tiny-voca-ru by default) or a local directory, automatically selecting PyTorch (model.safetensors + config.json) or ONNX Runtime (model.onnx + model.json).

Python API

from tiny_voca_ru import Synthesizer

# Load from Hugging Face Hub (defaults to "egnch/tiny-voca-ru")
synth = Synthesizer.from_pretrained("egnch/tiny-voca-ru")
# Explicit ONNX Runtime backend (downloads only onnx/model.onnx + onnx/model.json)
# synth = Synthesizer.from_pretrained("egnch/tiny-voca-ru", backend="onnx")
# Local safetensors or ONNX directory
# synth = Synthesizer("checkpoints/tiny-voca-ru")

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" | "ruaccent" | "none"
    pause_seconds=0.0,  # optional silence gap between sentences
)

CLI

# Uses Hugging Face Hub ("egnch/tiny-voca-ru") by default
tiny-voca-ru-speak \
  --voice F \
  --text "Привет! Билет стоит 1250 рублей." \
  --output out.wav

# Explicit ONNX Runtime backend
tiny-voca-ru-speak \
  --model egnch/tiny-voca-ru \
  --backend onnx \
  --voice M \
  --text "Привет! Билет стоит 1250 рублей." \
  --output out_onnx.wav

# Local checkpoint or ONNX directory
tiny-voca-ru-speak \
  --model checkpoints/tiny-voca-ru \
  --voice F \
  --text "Привет! Билет стоит 1250 рублей." \
  --output out_local.wav

Russian Text Frontend

tiny_voca_ru.frontend has no PyTorch or eSpeak dependency. It normalizes Russian text (numbers, dates, times, currencies, units, Roman numerals, abbreviations, acronyms, Latin words), places lexical stress, and outputs a 61-symbol IPA phoneme sequence.

from tiny_voca_ru import accentuate, normalize, phonemize, transcribe

normalize("В 1999 году билет стоил 12,50 руб.")
accentuate("Я открыл зам+ок ключом.")  # explicit '+' marks are preserved
transcribe("Здравствуйте!")  # readable IPA with ˈ
seq = phonemize("Привет, мир!", stress="lite")
Stress Provider Dependencies Notes
lite (default) Bundled stress.txt.gz + accent.onnx (onnxruntime) Fast, offline, runs in browser via JS accent.onnx
ruaccent tiny-voca-ru[ruaccent] Full ruaccent (turbo3.1) neural homograph resolver
none None Preserves existing + stress marks only

Custom-Voice Fine-Tuning

1. Prepare a dataset

Create a metadata.csv (or metadata.jsonl) with columns audio, text, and optional speaker, split, id:

audio,text,speaker
wavs/001.wav,Добрый день! Как ваши дела?,alice
wavs/002.wav,Сегодня отличная погода.,alice

Extract 24 kHz features and wav2vec2 CTC alignments:

tiny-voca-ru-prepare \
  --metadata data/metadata.csv \
  --audio-root data \
  --output work/alice-prepared \
  --val-ratio 0.05

2. Fine-tune from a pretrained checkpoint

When fine-tuning onto new speakers, tiny-voca-ru-train copies compatible weights and initializes new speaker embeddings from --init-voice (or the mean of pretrained voices):

tiny-voca-ru-train \
  --data-dir work/alice-prepared \
  --init-from checkpoints/tiny-voca-ru \
  --init-voice F \
  --output work/alice-run \
  --steps 5000 \
  --learning-rate 1e-4 \
  --save-every 500

Add --freeze-encoder for small datasets (< 30 min). Resume with --resume work/alice-run/last.pt. Safetensors checkpoints are saved to work/alice-run/inference/ and work/alice-run/best/.

ONNX Export & Legacy Checkpoint Conversion

# Export safetensors checkpoint -> model.onnx + model.json
tiny-voca-ru-export \
  --checkpoint checkpoints/tiny-voca-ru \
  --output checkpoints/tiny-voca-ru-onnx

# Convert legacy .pt checkpoint -> model.safetensors + config.json
tiny-voca-ru-convert \
  --source path/to/tiny-voca-ru.pt \
  --output checkpoints/tiny-voca-ru

Browser Demo

The static demo in web/public/ runs tiny_voca_ru.frontend in Pyodide and model.onnx + accent.onnx via ONNX Runtime Web entirely offline:

bash web/build_assets.sh checkpoints/tiny-voca-ru-onnx
python3 -m http.server --directory web/public 8000

Development

uv sync --python 3.12 --extra torch --extra onnx --extra train --group dev
uv run pytest -q
uv run ruff check .
uv run ruff format --check .
uv run basedpyright

Citation

Machine-readable citation metadata is available in CITATION.cff.

@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}
}

Please also consider citing the original tiny-voca by Muhammed Nazeem (2026).

License & Attribution

Code and weights are licensed under the Apache License 2.0 (LICENSE). See NOTICE for attribution to tiny-voca (Muhammed Nazeem, Apache-2.0) and ruaccent (MIT License, bundled stress.txt.gz and accent.onnx in src/tiny_voca_ru/frontend/data/).

Metadata

Release files for tiny-voca-ru 0.1.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for tiny-voca-ru 0.1.0
File Size Uploaded Provenance Actions
tiny_voca_ru-0.1.0.tar.gz 2.0 MB Without provenance Details for tiny_voca_ru-0.1.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for tiny-voca-ru 0.1.0
File Interpreter ABI Platform Actions
tiny_voca_ru-0.1.0-py3-none-any.whl Python 3 none any Details for tiny_voca_ru-0.1.0-py3-none-any.whl

Total release size: 4.0 MB

Release files / tiny_voca_ru-0.1.0.tar.gz

Download URL tiny_voca_ru-0.1.0.tar.gz
Size 2.0 MB
Tags Source
SHA-256 checksum
How to use checksums
6d402ae152d597a9577bcf5ed34d11a9a3a9516f18be9b7f3a82083b2a76203b
BLAKE2b-256 checksum
How to use checksums
f1dd9ed687a9aebd7d93e214181d59d954ac5c1a580a8576c0129834de692bd4
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.13.9 {"installer":{"name":"uv","version":"0.13.9","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Debian GNU/Linux","version":"13","id":"trixie","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

Release files / tiny_voca_ru-0.1.0-py3-none-any.whl

Download URL tiny_voca_ru-0.1.0-py3-none-any.whl
Size 2.0 MB
Tags Python 3
SHA-256 checksum
How to use checksums
69c7f6bd059f6b275445f207feb6bcf570f0c94fc38590b3699f371db6cb8bbf
BLAKE2b-256 checksum
How to use checksums
ce29af6fa9a47b1447bd4f4b43a2346e488bb41d5a97f5b9857c811c75f2e014
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.13.9 {"installer":{"name":"uv","version":"0.13.9","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Debian GNU/Linux","version":"13","id":"trixie","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

Security for tiny-voca-ru 0.1.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.1.0 This release

2 release files No provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page