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)
ReleaseInsight · On RadarHabr AI→ original

Russian Voice AI Squeezed Into 21 MB: tiny-voca-ru Runs Offline on a CPU

The author adapted the open tiny-voca speech synthesis model to Russian. It weighs about 21 MB, has 4.5M parameters and runs offline on an ordinary processor. Training used 12 hours of studio speech from the Dialogs dataset, and RUAccent places stress marks, with eSpeak as a fallback. For the demo, the author built a bot: microphone, speech recognition, a local qwen3.5:9b and voice output. For short replies, notifications and phone bots, that is enough without a cloud subscription.

AI-processed from Habr AI; edited by Hamidun News
Russian Voice AI Squeezed Into 21 MB: tiny-voca-ru Runs Offline on a CPU
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Full 100,000-step runabout 12.5 hours on a V100 versus 35–38 on a P40
Key metric #2A 125 W limit on the V100 gave 2.26 steps/s instead of 1.66 at 225 W
Key metric #3Two voices, Apache-2.0 license, but the data falls under OpenRAIL: read the terms before commercial use

Voice assistants usually run into the same wall: high-quality speech synthesis demands hundreds of megabytes of weights, a GPU or a cloud API. That means paying for traffic, depending on someone else's service and asking where users' voices end up. The developer behind tiny-voca-ru went the opposite way: they took a tiny open model and taught it to speak Russian. The result is about 21 MB in ONNX format and 4.5M parameters. It runs offline on an ordinary CPU.

What exactly was done

The base is the open tiny-voca model. It had no Russian profile, so the author added one and trained the model from scratch on Russian data. The training set was the Dialogs dataset: studio-recorded emotional dialogues by actors. From it, the author picked two voices, a female one (M1) and a male one (D1). The initial pool was 17.42 hours; after filtering by alignment, labeling confidence and unknown phonemes, about 12 hours of clean speech remained. A spot check of 60 clips through speech recognition gave a word error rate of 3.3%. That does not guarantee flawless data, but it suggests the labeling is not falling apart.

A separate piece of work is the Russian "frontend", the step that turns text into sounds. Phonemes come from eSpeak NG. Stress marks are taken from Dialogs for phrases that appear in the corpus. For arbitrary text, RUAccent places them and also resolves homographs. It is not a single point of failure: if stress placement is unavailable, the system falls back to eSpeak. There is also a small exceptions dictionary for service names and words whose pronunciation must stay stable. Text is phonemized in whole sentences, otherwise the preposition "s" ("с") may be pronounced as a standalone letter.

Hardware and an unexpected power-limit effect

A 15,000-step pilot took about 4.7 hours on a Tesla P40. Even at that stage, Whisper large-v3 recognized the control phrases without errors, although the intonation was still flat. A full 100,000-step run on the P40 would have taken 35–38 hours, while on a Tesla V100 32 GB it finished in roughly 12.5 hours of GPU time. The speedup was 4–5x in the pre-GAN stage and about 2.4x in the GAN phase.

One curious observation: on the V100, capping power at 125 W gave 2.26 steps per second at 70 °C, while at 225 W the speed was 1.66 steps/s at 74 °C. The author explains this by saying the card ran more steadily at the lower limit and data preparation kept up with it. The author himself cautions that this is not a universal law: in most workloads, lowering power reduces speed. The practical takeaway is more modest: if training is bottlenecked by the data loader, it is worth measuring a few power limits, since it is a free test.

The demo assistant and its limitations

The model is shown as more than a "speak this text" button. At voice.lmchat.ru you can type text, pick a voice and speed, exchange voice messages with a bot or call it from the browser. The pipeline is: microphone, speech recognition, a local qwen3.5:9b LLM via Ollama, the Russian frontend, synthesis, playback. For real telephony, the setup uses Asterisk and AudioSocket: the bot picks up the call, and tiny-voca voices the reply.

The limitations are stated honestly. Quality is below large cloud TTS systems, especially on complex intonation and long texts. Stress is sometimes misplaced in rare words. The model is tuned for short replies, not audiobooks. The pilot metrics (pitch error of 3.07 semitones, pause accuracy of about 92%) show that the speech is intelligible, but WER on control phrases does not measure naturalness.

There is also a legal nuance. The code and model are under Apache-2.0, but the training data is distributed under OpenRAIL, and the voices belong to real actors. For commercial projects, the terms need to be checked separately, and you should not advertise using these voices without the rights.

For practitioners, the main takeaway is this: a voice bot for notifications, kiosks, call scenarios or embedded devices no longer requires cloud TTS. Pairing a small synthesis model with a local 9B-parameter LLM removes recurring API payments and keeps data inside your own perimeter. Do not expect perfect speech, but the barrier to entry drops sharply.

Story

Key Questions Answered

What exactly was done?

The base is the open tiny-voca model. It had no Russian profile, so the author added one and trained the model from scratch on Russian data. The training set was the Dialogs dataset: studio-recorded emotional dialogues by actors. From it, the author picked two voices, a female one (M1) and a male one (D1). The initial pool was 17.42 hours; after filtering by alignment, labeling confidence and unknown phonemes, about 12 hours of clean speech remained.

What are the key details on Hardware and an unexpected power-limit effect?

A 15,000-step pilot took about 4.7 hours on a Tesla P40. Even at that stage, Whisper large-v3 recognized the control phrases without errors, although the intonation was still flat. A full 100,000-step run on the P40 would have taken 35–38 hours, while on a Tesla V100 32 GB it finished in roughly 12.5 hours of GPU time. The speedup was 4–5x in the pre-GAN stage and about 2.4x in the GAN phase.

What are the key details on The demo assistant and its limitations?

The model is shown as more than a "speak this text" button. At voice.lmchat.ru you can type text, pick a voice and speed, exchange voice messages with a bot or call it from the browser. The pipeline is: microphone, speech recognition, a local qwen3.5:9b LLM via Ollama, the Russian frontend, synthesis, playback. For real telephony, the setup uses Asterisk and AudioSocket: the bot picks up the call, and tiny-voca voices the reply. The limitations are stated honestly.

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…