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

GigaChat Pro Marked Stress in 95.8% of Words, but 32 Errors Surfaced Only in Small Batches

Russian TTS engines put stress in the wrong place: «гудни» (gudni) is read as «гУдни», and the pronunciation dictionaries in ElevenLabs are locked on the author's plan. He gave GigaChat Pro 7789 words and got back a lexicon of 7462 words, or 95.8% coverage. But for 32 words, such as «проскочила» (proskochila) and «куплена» (kuplena), the model placed the stress differently from run to run, and only small-batch rechecks caught it. The takeaway: a bulk run is just a list of candidates, and a human should decide the disputed cases.

AI-processed from Habr AI; edited by Hamidun News
GigaChat Pro Marked Stress in 95.8% of Words, but 32 Errors Surfaced Only in Small Batches
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The substitution dictionary grew from 51 to 101 words, and none of the new entries could have been guessed by hand
GigaAM v3 checks the voice-over222 MB of weights, MIT license, files up to ~50 seconds
Key metric #3The Lite model produces junk for stress marking; only GigaChat-2-Pro works

Russian speech synthesizers (TTS) regularly put stress in the wrong place: the word «гудни» (gudni) is read as «гУдни», and «проскочила» (proskochila) and «куплена» (kuplena) come out with errors. The author of this write-up ran into the problem in three engines at once. Pronunciation dictionaries in ElevenLabs are locked on his plan, so he kept a manual substitution dictionary of 51 words.

Then he had GigaChat Pro mark stress across his entire bank of lines: 7789 words became a lexicon of 7462 words, or 95.8% coverage. That sounds like a win, but the main lesson lies elsewhere: 32 words in the list turned out to be unreliable, and the only way to see it was to recheck in small batches.

For anyone labeling data in bulk with an LLM, this is a ready-made practical rule.

GigaChat as a stress-dictionary builder

The idea is simple: hand the language model the whole text bank and build the stress lexicon automatically instead of typing in words by hand. GigaChat has a free tier for a year: 250 million tokens for Lite, 40 million for Pro, 25 million for Max and 50 million for Ultra. Registration uses a Client ID with the scope GIGACHAT_API_PERS, and connecting takes a single line of code that specifies the GigaChat-2-Pro model.

The choice of model turned out to matter. Lite is useless for this task: it "produces junk," according to the author. Only Pro works.

There is also a minor quirk. GigaChat places the stress mark after the stressed vowel rather than before it, as the ruaccent library does, and sometimes cuts off the end of a word. Fixing the format in the prompt and patching the parser solves this.

The main lesson: large batches lie with confidence

The finished lexicon cannot be imported into the dictionary directly. When the author rechecked the words in small batches, he found 32 tricky cases, including «проскочила» and «куплена»: in one run the model put the stress one way, in another run a different way. In a large batch such discrepancies get lost, and the answer looks equally confident.

This produced a rule that now applies permanently in the project:

  • a bulk run is only a source of candidates;
  • every candidate goes through a small-batch check;
  • a human makes the final decision on discrepancies between runs.

The result: the dictionary grew from 51 to 101 words, and the author would not have guessed any of the 50 new words himself. The rule is not limited to stress marks. If you ask an LLM to assign tags, categories or normalize data into a uniform format, the same thing can happen to you: across thousands of items, the instability is invisible. A cheap safeguard is to rerun the disputed items in small batches and compare the answers.

GigaAM as a quality controller for voice-over

The second task is to check the finished voice-over: transcribe the generated wav file and compare the words with the reference text. GigaAM v3 via sherpa-onnx fit the job: the weights take up 222 MB, and the license is MIT. On clean, unprocessed tracks it is fast and accurate. As a bonus, the author ran a six-minute dictaphone recording through it and got a nearly error-free transcript in seconds.

The limitations are real, too:

  • on voice-over with a radio effect (interference and a compressor on "Kanal 15"), GigaAM gets confused, and whisper stays in place there;
  • it does not accept long files whole: the practical limit is about 50 seconds, so tracks are cut into 30-second windows;
  • GigaChat's Lite model remains useless for stress marking.

Both tools are built into the working protocol: the script stress_suggest.py with the --giga flag proposes stress marks, while qa_recent.py and qa_recast.py with the --engine gigaam parameter check fresh resyntheses.

What to take away

For anyone building Russian voice-over or voice agents, the setup is reproducible in an evening. The free GigaChat Pro tier collects candidates, small batches filter out unstable answers, and a human looks only at the discrepancies. GigaAM works as an inexpensive automatic check on clean audio, but for tracks with effects it is better to keep whisper. The author sums it up without pathos: a neural network is a tool like a screwdriver, only more expensive. And one telling detail: the model never managed to place the stress in «гудни» confidently, not even once.

Story

Key Questions Answered

What are the key details on GigaChat as a stress-dictionary builder?

The idea is simple: hand the language model the whole text bank and build the stress lexicon automatically instead of typing in words by hand. GigaChat has a free tier for a year: 250 million tokens for Lite, 40 million for Pro, 25 million for Max and 50 million for Ultra. Registration uses a Client ID with the scope GIGACHAT_API_PERS, and connecting takes a single line of code that specifies the GigaChat-2-Pro model. The choice of model turned out to matter.

What are the key details on The main lesson: large batches lie with confidence?

The finished lexicon cannot be imported into the dictionary directly. When the author rechecked the words in small batches, he found 32 tricky cases, including «проскочила» and «куплена»: in one run the model put the stress one way, in another run a different way. In a large batch such discrepancies get lost, and the answer looks equally confident.

What are the key details on GigaAM as a quality controller for voice-over?

The second task is to check the finished voice-over: transcribe the generated wav file and compare the words with the reference text. GigaAM v3 via sherpa-onnx fit the job: the weights take up 222 MB, and the license is MIT. On clean, unprocessed tracks it is fast and accurate. As a bonus, the author ran a six-minute dictaphone recording through it and got a nearly error-free transcript in seconds.

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…