---
title: xVibePocketTTS
emoji: 🪆
colorFrom: red
colorTo: blue
sdk: gradio
sdk_version: 6.10.0
app_file: app.py
pinned: false
models:
- ruaccent/accentuator
---

# xVibePocketTTS — ZeroGPU demo

Public demo of the Russian Pocket TTS step80000 checkpoint. The Space provides three built-in
voices and optional zero-shot voice cloning from an uploaded recording. RUAccent runs before
synthesis. When a reference is uploaded or recorded, it takes priority over the selected built-in
voice.

Telegram автора:
<a href="https://t.me/xVibeNot" target="_blank" rel="noopener noreferrer">t.me/xVibeNot</a>

The model weights, tokenizer, and built-in model-specific voice states are stored in the private
`Brakanier/pocket-tts-ru-private` model repository. The public Space accesses them through its
`HF_TOKEN` secret. Source WAV recordings are not included, exposed in the UI, or needed at runtime.

The automatic accentizer runs on CPU. The UI shows stress as a plus before the stressed vowel
(`з+амок`, `зам+ок`), and users can use the same notation to override RUAccent. Immediately before
inference, the app converts these marks to the canonical combining acute form used during training.
Because `ё` is inherently stressed, the app removes redundant `+ё` and `ё́` marks.

## Public Space contents

The public repository contains the Gradio app, model architecture config, and vendored Python
wheels. It does not contain:

- `model/model.safetensors`
- `model/tokenizer.model`
- `model/tokenizer.vocab`
- `voices/*.safetensors`

At runtime, `private_assets.py` resolves missing assets to the private model repository. The
default revision is pinned to an immutable private-repository commit SHA.

## Deploy the public ZeroGPU Space

```bash
hf repos create Brakanier/xVibePocketTTS \
  --type space --sdk gradio --flavor zero-a10g --public

hf upload Brakanier/xVibePocketTTS ./hf_space \
  --type space \
  --exclude "model/model.safetensors" --exclude "model/tokenizer.*" \
  --exclude "voices/**" --exclude "ruaccent_data/**" --exclude "**/__pycache__/**" \
  --commit-message "Deploy public Pocket TTS RU ZeroGPU demo"
```

Add `HF_TOKEN` as a Space secret with read access to the private model repository.

## Local CPU launch

From this directory, after installing `requirements.txt`:

```bash
POCKET_TTS_DEVICE=cpu python app.py
```

The ZeroGPU deployment uses `POCKET_TTS_DEVICE=cuda` by default. Local launch automatically uses
the model files already present under `model/` and `voices/`. The vendored RUAccent wheel was built
from commit `3ac0ad5f6508f6c1a4a604220042232c70a7baf9`; it downloads only the resources needed by the
`turbo3.1` pipeline from the pinned model revision on first startup.