Silero TTS Server

Wyoming protocol TTS server wrapping Silero models. Supports Home Assistant integration, streaming audio, and optional per-voice LLM normalization.

Quick Start

Local Development

1. Clone the repository

   git clone https://github.com/dkindev/silero-server.git
   cd silero-server

2. Create a virtual environment

   python -m venv .venv
   source .venv/bin/activate

3. Install dependencies

   uv sync --extra dev

4. Run the application

   python -m src.main

   The server listens on tcp://127.0.0.1:10200 by default.

5. Test with HTTP server

   uv run --with "wyoming[http]" python -m wyoming.http.tts_server --uri tcp://127.0.0.1:10200

Docker

1. Build the image

   docker build -t silero-server:latest .

2. Run the container

   docker run -p 10200:10200 \
     -e TTS_ENV_TYPE=production \
     -v ./data:/app/data \
     -v ./models:/app/models \
     silero-server:latest

   The server exposes port 10200 for Wyoming TCP connections.

Configuration

Configuration is resolved with the following priority (highest first):

1. CLI arguments — --uri, --torch_device, --sample_rate, etc.
2. Environment variables — TTS_URI, TTS_TORCH__DEVICE, TTS_SAMPLE_RATE, etc.
3. config.yml — YAML file at the project root
4. Defaults — Default values in code

See config.yml for full configuration reference with all fields, CLI argument names, and environment variable names.

Voice and Model Configuration

Voice and model mappings are defined in data/data.yml. The file has two top-level sections: models and prompts.

Models

models:
  v5_5_ru:
    language: ru
    warmup: true
    hash_prefix: 50081637b602126ee06cb3bc8a744d25651d2da149ee8864b9a379bfdd934437
    locales:
      ru_RU:
        voices:
          - name: aidar
          - name: eugene
          - name: baya
          - name: kseniya
          - name: xenia
  v3_en:
    language: en
    locales:
      en_US:
        voices:
          - name: henry
            speaker: en_2
          - name: sophia
            speaker: en_3

Each model key (e.g., v5_5_ru) is the model name. Supported fields:

Field	Required	Default	Description
language	Yes	—	Silero language identifier (ru, en, etc.)
enabled	No	true	false to disable the model and exclude its voices
warmup	No	false	true to preload at startup; false for lazy loading
hash_prefix	No	—	SHA-256 hash prefix for download integrity verification
locales	Yes	—	Map of locale codes to voice lists

Adding a new model

1. Add a model entry under models: in data/data.yml with the Silero language identifier.
2. Add locales with voices referencing the new model.

Disabling a model

Set enabled: false on the model entry. Its voices are automatically excluded from discovery.

Voices

Each voice entry has the following fields:

Field	Required	Default	Description
name	Yes	—	Voice name
speaker	No	voice name	Silero's native speaker identifier
model	No	parent model	Explicit model reference (usually inherited from parent)
normalization.text.enabled	No	true	Override text normalization enabled state
normalization.text.type	No	simple	Override normalization type (simple, llm)
normalization.text.prompt_id	No	—	Reference a custom LLM prompt from prompts
normalization.ssml.enabled	No	false	Override SSML normalization enabled state
normalization.ssml.type	No	simple	Override SSML normalization type
normalization.ssml.prompt_id	No	—	Reference a custom LLM prompt for SSML

Voices are identified by their computed ID: {locale}-{model}-{name} (e.g., ru_RU-v5_5_ru-aidar).

Adding or renaming voices

Add or edit entries under locales.*.voices in data/data.yml. Set speaker to the Silero native speaker name only if it differs from the voice name.

Voice normalization

Per-voice override of the server-wide normalization settings (defined in config.yml). Override the normalization type, disable it entirely, or reference a custom LLM prompt.

models:
  v5_5_ru:
    locales:
      ru_RU:
        voices:
          - name: aidar
            normalization:
              text:
                enabled: true
                type: llm
                prompt_id: ru_text

prompts:
  - id: ru_text
    text: "Normalize Russian text for TTS..."
    model: qwen3.5:4b

Without a normalization block on a voice, the server-wide defaults from config.yml apply.

Prompts

Reusable LLM prompt definitions referenced by prompt_id in voice normalization.

Field	Required	Default	Description
id	Yes	—	Unique identifier referenced by prompt_id
text	Yes	—	LLM prompt instruction text
model	Yes	—	LLM model name (e.g., qwen3.5:4b)

Audio

Supported Input Types

• TEXT — plain text
• SSML — Speech Synthesis Markup Language

Audio Output

• Format: raw PCM (WAV-compatible)
• Sample rate: configurable (8000, 16000, 24000, 48000 Hz)
• Bit depth: 16-bit signed integer
• Channels: mono (1 channel)

Development

Running Tests

pytest tests/

Linting and Formatting

ruff check src tests
ruff format src tests

Pre-commit Hooks

pre-commit install

CI/CD

The project uses GitHub Actions for continuous integration. On every push:

1. Lint: Ruff checks code quality across Python 3.11, 3.12, 3.13, and 3.14
2. Test: pytest runs all tests across the same Python versions

See .github/workflows/ci.yml for details.

License

MIT License — see LICENSE file for details.