Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableHabr AI→ original

Glasvo voices 17 seconds of Russian speech in 0.08 s on a home RTX 3070

Tired of online voice-over services with character limits and subscriptions, a developer built Glasvo, a free open-source program that reads text aloud on a home PC. On an RTX 3070, the Silero engine voiced a 285-character paragraph (17 seconds of audio) in 0.08 s, and voice cloning took 4.4–5 s. For Russian text, the local stack can replace a subscription, but it needs about 10 GB of disk space and a license check.

AI-processed from Habr AI; edited by Hamidun News
Glasvo voices 17 seconds of Russian speech in 0.08 s on a home RTX 3070
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1Everything together weighs about 10 GB; cloning needs only 2 GB of VRAM
Key metric #2Silero silently skips digits and Latin letters, so they must be replaced with words beforehand
Key metric #3Silero and the Russian F5-TTS checkpoint are under CC BY-NC 4.0: read the terms before monetizing

If you regularly need to voice Russian text, you know the annoyance: the free character limit runs out, and then you are asked for a subscription. The author of Glasvo solved the problem differently and built a program that reads text aloud right on a home computer. The code is open under the MIT license.

Here is the speed: on a gaming RTX 3070 GPU (8 GB of memory), the Silero engine voiced a 285-character paragraph, just over 17 seconds of speech, in 0.08 seconds. Voice cloning, where the program reads text in a voice taken from a short sample, takes 4.

4–5 seconds on the same text. The practical takeaway: for Russian text, local voice-over is already good enough to replace a subscription, but it has its pitfalls, and the author has honestly written them down.

What's inside and how much it weighs

The tkinter window combines three engines: Silero, Microsoft Edge online voices, and voice cloning via XTTS-v2 and F5-TTS. The models themselves are called with roughly ten lines of code. Most of the work went into the details around them.

The program itself weighs about 8 MB; the installer downloads the rest. The Python 3.12 environment with PyTorch 2.5.1 (CUDA 12.1) and 168 packages takes about 5.5 GB. Silero models weigh from 39 to 141 MB each, XTTS-v2 is 2.0 GB, the Russian F5-TTS checkpoint is 1.4 GB, and Whisper small, which transcribes the voice sample, is 0.9 GB. The total is about 10 GB.

A GPU is needed only for voice cloning. Silero and Edge work without one. For XTTS and F5, the measured peak VRAM usage was about 2 GB, but the author recommends a card with at least 6 GB.

Measurements were taken on a single 285-character Russian text after warm-up:

  • Silero v5_ru on GPU: 0.08 s, on CPU: 0.53 s;
  • XTTS-v2: 5 s for 18.7 s of audio, model load 13 s;
  • F5-TTS: 4.4 s for 17.3 s of audio, load 4 s.

The author notes that these are single runs on one machine, not a benchmark. The order of magnitude is still clear: Silero runs hundreds of times faster than real time, and voice cloning about three to four times faster.

What Silero silently drops

The most unpleasant surprise: Silero does not raise an error on digits and Latin letters; it simply skips them. Instead of "dve tysyachi dvadtsat shest" ("two thousand twenty-six"), the audio has a gap. English words in the Russian models are cut out just as quietly.

That is why two functions sit in front of the model. The first converts numbers into words via num2words and handles fractions, lists and minus signs. It is not perfect with grammatical cases: "v 1990 godu" ("in 1990") may sound wrong, and the author advises writing such numbers out as words. The second function handles Latin text in three steps:

  1. Common words are taken from a dictionary: YouTube becomes "yutub", Windows becomes "vindous".
  2. Words in capital letters with no vowels are spelled out letter by letter: USB becomes "yu-es-bi".
  3. Everything else is transliterated.

The check was done the engineering way: phrases with English words were voiced and then transcribed back through Whisper, and all the words were found. Cloning was verified in a closed loop: a Silero voice was used as the sample, XTTS and F5 read Russian text in it, and Whisper transcribed the result. The words matched 100% for both engines.

There are also length limits. A chunk of text for Silero should not exceed about 700 characters, for XTTS-v2 about 182, and for F5 the author chose 400. Long text is split by sentence and stitched back together. The Russian Silero models have only five voices; another 29 sit in v5_cis_base and speak with an accent.

Dependencies, installation and licenses

The author spent the most time fighting package versions. Cloning pulls in coqui-tts, f5-tts and transformers, and the package manager runs into conflicts. torch had to be pinned to 2.5.1: transformers 5 and diffusers 0.43 require version 2.6 or higher, which breaks everything else. The setup also needed transformers below version 5, huggingface_hub below 1.0, and omegaconf 2.3.1 for hydra, which F5 requires. requirements.txt contains an exact copy of the working environment (pip freeze), not version ranges.

The installation is designed for people without Python: a script downloads uv, uses it to install Python 3.12, and creates a venv. Without an NVIDIA GPU, the CPU build of PyTorch is used. There are two traps: the uv cache must be kept in a folder with a short path, otherwise the antlr4 package fails to build, and models must not be downloaded into the same folder by two processes at once.

The main limitation is licensing. The code is MIT, but the models are mostly non-commercial: Silero and the Russian F5-TTS checkpoint are released under CC BY-NC 4.0, XTTS-v2 under the Coqui Public Model License, and only Whisper is MIT. For monetized videos, you need to read the terms of the specific model separately. You should clone only your own voice or the voice of a person who has given consent.

The bottom line for practitioners: local Russian voice-over can be put together in an evening. Set aside 10 GB of disk space, add preprocessing for digits and Latin text, pin package versions, and check licenses right away if you plan commercial use.

Story

Key Questions Answered

What's inside and how much it weighs?

The tkinter window combines three engines: Silero, Microsoft Edge online voices, and voice cloning via XTTS-v2 and F5-TTS. The models themselves are called with roughly ten lines of code. Most of the work went into the details around them. The program itself weighs about 8 MB; the installer downloads the rest. The Python 3.12 environment with PyTorch 2.5.1 (CUDA 12.1) and 168 packages takes about 5.5 GB.

What Silero silently drops?

The most unpleasant surprise: Silero does not raise an error on digits and Latin letters; it simply skips them. Instead of "dve tysyachi dvadtsat shest" ("two thousand twenty-six"), the audio has a gap. English words in the Russian models are cut out just as quietly. That is why two functions sit in front of the model. The first converts numbers into words via num2words and handles fractions, lists and minus signs.

What are the key details on Dependencies, installation and licenses?

The author spent the most time fighting package versions. Cloning pulls in coqui-tts, f5-tts and transformers, and the package manager runs into conflicts. torch had to be pinned to 2.5.1: transformers 5 and diffusers 0.43 require version 2.6 or higher, which breaks everything else. The setup also needed transformers below version 5, huggingface_hub below 1.0, and omegaconf 2.3.1 for hydra, which F5 requires. requirements.txt contains an exact copy of the working environment (pip freeze), not version ranges.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…