Technical Context
I saw a brief but telling piece of feedback: in Claude Code, Russian voice input proved weak. For me, this isn't just about one failed test—it's about an old AI implementation problem: a language is supposedly supported, yet real-world workflows still fall apart.
I dug into the docs. Claude Code does offer /voice; you can set the dictation language in settings or via a BCP 47 code, and Russian is listed as supported. So on paper, everything looks great: the feature exists, the language is there, the barrier to entry is low.
But here's what I couldn't find: any solid public metrics on Russian STT quality. No word error rate (WER), no breakdown by noise, accents, mixed speech, or dev terminology. Without those, any language support looks like a checkbox in the UI, not an engineering promise.
And the most interesting part: voice input for code lives under harsher conditions than regular transcription. It has to accurately catch English tokens inside Russian phrases—filenames, commands, symbols, API snippets, sometimes numbers and punctuation. If the engine stumbles on the ru/en switch, using it for more than a couple of minutes becomes painful.
I wouldn't be surprised if it feels fine in English and far worse in Russian. That's the classic story for multilingual products: marketing says "20 languages," but real usefulness is wildly uneven.
What This Means for Automation
For a developer, it's just annoying. For a business, the consequences are harder: voice-based AI automation stops being a reliable input channel when an employee speaks Russian and mixes in commands, entity names, and technical terms.
The teams that lose are those hoping to quickly bolt on voice-first scenarios in IDEs, support, or internal tools. The ones winning right now are those building a pipeline through a separate, strong STT layer and only then feeding text to an LLM.
I wouldn't currently bank on Claude Code voice as the primary interface for Russian-language processes without testing it on your own data. At Nahornyi AI Lab, we specialize in exactly these spots: figuring out where a simple AI integration will suffice and where you need a different STT contour or a custom AI solutions architecture so that automation doesn't break down on the very first call.