Silero Models

0.5.5 · active · verified Fri Apr 17

Silero Models provides a collection of pre-trained enterprise-grade Text-to-Speech (TTS) models primarily focused on Russian and CIS languages, as well as speech-to-text models. It leverages PyTorch for model inference, offering high-quality and fast speech generation. The library is actively maintained with frequent updates, currently at version 0.5.5, with a focus on expanding language support and model quality.

Common errors

```
ModuleNotFoundError: No module named 'torchaudio'
```
cause The `torchaudio` library, which is critical for Silero's audio processing, is not installed.

fix Install `torchaudio`: `pip install torchaudio`.
```
RuntimeError: Requested model is not available for language 'xx'. Available languages: [...]
```
cause The specified language or model ID in `torch.hub.load` is either incorrect or not supported by the current Silero model version.

fix Check the available languages and model IDs in the official Silero Models repository's `models.yml` file or the quickstart guide. Ensure you are using a currently supported model, especially after v5.0 breaking changes.
```
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
```
cause A common PyTorch error indicating that tensors involved in an operation are not on the same computational device (e.g., one on GPU, one on CPU).

fix Ensure all tensors, including the model itself, input text embeddings (if manually processed), and any other relevant data, are moved to the same device (e.g., `model.to(device)` and `input_tensor.to(device)`).

Warnings

breaking Silero v5.0 and later versions (e.g., v5.2) deprecated and removed legacy models (v1 and v2) and tools. Attempting to load these older models will result in errors.
Fix: Migrate to using `v5_ru` or `v5_cis_*` model identifiers. Refer to the official GitHub repository's `models.yml` or quickstart for current supported model IDs and languages.
breaking As of v5.1, `torchaudio` was removed as a direct dependency of the `silero` pip package. While `silero` installs without it, `torchaudio` is still essential for most functionalities (e.g., audio I/O, many model operations).
Fix: Always install `torchaudio` explicitly alongside `silero` using `pip install torchaudio` or include it in your requirements file.
gotcha The core dependencies `torch`, `torchaudio`, and `soundfile` are not always automatically installed by `pip install silero`. Missing these will lead to `ModuleNotFoundError` or `RuntimeError` during model loading or inference.
Fix: Always install all necessary dependencies explicitly: `pip install torch torchaudio silero soundfile`.
gotcha The license for Silero Models changed to GNU AGPL 3.0 in v5.4. Previous versions used CC BY-NC 4.0. Users should be aware of the implications of the AGPL-3.0 license for commercial or proprietary use.
Fix: Review the GNU AGPL 3.0 license agreement to ensure compliance with its terms, especially if using Silero Models in a commercial or distribution context.

Install

pip install torch torchaudio silero soundfile Recommended installation with core dependencies

Imports

torch
```
import torch
```
torchaudio
```
import torchaudio
```
silero.utils.save_audio
```
from silero.utils import save_audio
```
silero.utils.read_audio
```
from silero.utils import read_audio
```

Quickstart

This quickstart demonstrates how to load a pre-trained Silero TTS model using `torch.hub.load` and synthesize speech from text. It uses the 'v5_ru' Russian model with a specific speaker. Make sure `torch`, `torchaudio`, and `soundfile` are installed as they are crucial prerequisites, even if not direct dependencies of the `silero` PyPI package itself.

import torch
import torchaudio

# Ensure PyTorch is available and get device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# --- TTS Quickstart ---
# Define model parameters
model_id = 'v5_ru'
language = 'ru'
speaker = 'baya'
put_accent = True
put_yo = True
sample_rate = 48000 # or 24000, 16000
text = 'В недрах тундры выдры в гетрах тырят в вёдра ядра кедров.'

try:
    # Load the Silero TTS model from torch.hub
    model, _ = torch.hub.load(repo_or_dir='snakers4/silero-models',
                              model='silero_tts',
                              language=language,
                              speaker=model_id,
                              put_accent=put_accent,
                              put_yo=put_yo)
    model.to(device)

    # Synthesize audio
    audio_tensor = model(text=text, speaker=speaker, sample_rate=sample_rate)
    
    # Example of saving audio (requires 'soundfile')
    # from silero.utils import save_audio
    # output_path = 'output_audio.wav'
    # save_audio(audio_tensor.cpu(), output_path, sample_rate)
    # print(f'Audio saved to {output_path}')
    
    print(f"Successfully synthesized audio. Tensor shape: {audio_tensor.shape}, Sample Rate: {sample_rate}")
except Exception as e:
    print(f"An error occurred during TTS synthesis: {e}")
    print("Please ensure PyTorch, TorchAudio, and potentially SoundFile are installed.")

view raw JSON →