Russian Emotional TTS

Russian Emotional TTS

Speaks Russian text with emotion and marked word stress, and can copy a voice from a 3 to 10 second clip. Clone only voices you have the right to use.

Made by Andreevai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a Russian text to speech and voice cloning model built on the F5-TTS design, which uses a diffusion transformer with flow matching. Andreevai fine-tuned it on expressive theatrical dialogues that carry stress annotations. The whole package comes to about 350 million parameters and outputs 24,000 Hz mono audio.

Who it suits: It suits creators and developers who need expressive spoken Russian, and anyone who wants correct stress on each word. The record lists Russian only.

Industry | Who uses it | How they use it | Example
Media | Video and audio creators | Narrate Russian scripts with emotion | A voice-over for a short video
Games | Indie developers | Give characters spoken Russian lines | Lines for a story character
Education | E-learning teams | Read lessons aloud with correct word stress | A lesson with stress marks on tricky words

People who need languages other than Russian can skip it, since the model was fine-tuned on Russian only.

Where to use it:

• Python, using the f5-tts package
• Google Colab, through the notebook the card offers, on a free T4 GPU
• A Gradio web page that the Colab notebook opens
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Russian narration | Creators | A wav file of the script read aloud
Voice matching | Developers | Speech in the timbre of a 3 to 10 second reference clip
Stress control | Teachers | Words pronounced with the stress you mark

How to set it up: You need a few things first.

• Python and pip
• A clean reference clip of 3 to 10 seconds, only of a voice you have the right to use
• The model.safetensors and vocab.txt files from this repository

1. Install the f5-tts package with the other two libraries.
2. Download the repository files.
3. Put your reference clip next to the script.
4. Run the Python code below.

pip install f5-tts soundfile torchaudio

pip install -U huggingface_hub
hf download Andreevai/f5-tts-russian-emotional

How to use it: You give the model a reference clip, its transcript and the text to speak. Put a plus sign before each stressed vowel to control the stress. The script picks the GPU when one is available and otherwise falls back to the CPU.

import torch
import soundfile as sf
from f5_tts.infer.utils_infer import infer_process, load_model, load_vocoder
from f5_tts.model import DiT

device = "cuda" if torch.cuda.is_available() else "cpu"

# Load neural vocoder
vocoder = load_vocoder(vocoder_name="vocos", device=device)

# Load acoustic DiT model
model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
model = load_model(
    DiT,
    model_cfg,
    "model.safetensors",
    vocab_file="vocab.txt",
    device=device
)

# Zero-shot inference
wave, sr, _ = infer_process(
    ref_audio="reference.wav",
    ref_text="Текст из образца голоса",
    gen_text="Прив+ет! Это русская нейрос+еть для озв+учки.",
    model_obj=model,
    vocoder=vocoder,
    device=device,
    nfe_step=32,
    speed=1.0
)

sf.write("output.wav", wave, sr)

Tool | What it does
load_vocoder | Loads the Vocos neural vocoder
load_model with DiT | Loads the acoustic model from model.safetensors and vocab.txt
infer_process | Generates speech from the reference clip, its text and the text to speak
Plus sign before a vowel | Marks the stressed vowel in the text
sf.write | Saves the audio as a wav file

Example: You supply a short clean reference clip with its transcript, then ask for a Russian greeting that introduces the tool, with plus signs on the stressed vowels. The script writes output.wav in the voice of the clip.

For newcomers and experts: A text to speech model reads written text aloud, and voice cloning copies how a speaker sounds. If you are new, run the Colab notebook, upload a short clip of a voice you may use and type a Russian sentence.

The model is the F5-TTS DiT Base v2 design, fine-tuned on more than 14 hours of emotional actor dialogues. It pairs a 337M diffusion transformer with a 13.5M Vocos vocoder at 24 kHz. The weights are plain safetensors, so loading avoids pickle files. The example runs 32 inference steps at normal speed.

Reference clip and its text go in
Text with stress marks is encoded
Diffusion transformer builds the speech
Vocos vocoder turns it into audio
Wav file is written to disk

Good to know:

• License: MIT, which the card describes as free for commercial and personal use.
• Cost to run: no API key or paid service named. A GPU helps, and the free Colab T4 works.
• Data it reaches: the reference audio and text you give it, and the wav file it writes. Use only voices you have the right to use.
• Not for: languages other than Russian, or cloning a voice without permission.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: f5-tts
• Parameters: 337M
• Weights: about 1.3 GB in F32; about 169 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 813
• Likes: 5
• License: mit
• Languages: ru
• Tags: f5-tts, safetensors, text-to-speech, speech, flow-matching, audio, russian, voice-cloning

Links

• Model: https://huggingface.co/Andreevai/f5-tts-russian-emotional
• Files: https://huggingface.co/Andreevai/f5-tts-russian-emotional/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Andreevai/f5-tts-russian-emotional

Then follow the usage in its model card. Nothing in this kit runs by itself.