Yandex SpeechKit

yandex's hosted speech API, and the best-measured Russian TTS in the only independent Russian evaluation this spoke can find — CER 0.6 on both tested voices in russian-tts-evaluation, lowest of twenty-one rows, with UTMOS 3.413 and 3.482 and the fastest cloud synthesis rate measured (xRT 0.028).

It is closed, hosted, and the first Russian-market commercial voice provider to get a page here; elevenlabs has held that slot for the English-speaking market since the spoke was founded.

The Russian voice roster

Thirteen Russian voices plus five Central-Asian-accented ones. The API has two generations and most voices are v3-only.

The role axis has no counterpart in the open field. Silero exposes speaker, rate, pitch and SSML; the emotional register is fixed. Here nine of the thirteen Russian voices carry two or three named deliveries — good, evil, strict, friendly, whisper — selectable per request. That is a coarser control than ElevenLabs' expressive mode and a far cheaper one, and it is the axis the open Russian models have nothing to say about.

The docs do not state sample rates per voice.

Price, which is the surprising part

From the published pricing policy (USD, excluding VAT; rouble pricing applies to Yandex.Cloud LLC customers and is not listed on the English page):

Roughly $11 per million characters. For a 500,000-character book — a 300-page novel, our estimate — that is about $5.50 of synthesis. elevenlabs' Multilingual v2 bills 1 credit per character, so the same book needs 500,000 credits: the $11/month Creator tier supplies 121,000 and the $99/month Pro tier supplies 600,000. Same book, ~$5.50 against ~$99, and the cheaper one has the better independently-measured clarity in Russian.

The v1/v3 split matters for that arithmetic. v3 bills per request, not per character, so the per-book cost depends on how you chunk the text — and v3 is where all the newer voices and most of the emotional roles live. The characters-per-request ceiling is not on the pricing page.

What is not established

• No independent measurement since 2024-07-12. The CER 0.6 figure is from russian-tts-evaluation and covers alexander and marina only; the other sixteen voices have never been measured by anyone outside Yandex. (Re-checked 2026-09-17: still true; no MOS or CER study of SpeechKit since.) (Re-checked 2026-09-28: still true for the voices. RuASD (arXiv 2604.02374) measured 37 Russian TTS systems and SpeechKit is not one of them. The recognition side has been measured: a Habr roundup by bobastia (10 February 2026, habr.com/ru/articles/993786) ran twelve speech-to-text APIs on 20 of its own YouTube clips and put SpeechKit last, WER 0.657 overall and 0.550 on the Russian-only subset. That is one hobbyist test against uploader subtitles, and it says nothing about the voices this page covers.)
• Nothing published about the front end. Whatever SpeechKit does about stress, ё and homographs — and its CER says it does something well — is undocumented. ruaccent and silero-stress at least publish their mechanisms; this is the same stage as a black box.
• Access is a real constraint, not a footnote. USD billing runs through entities in Serbia and Dubai, and payment and availability outside Russia have moved repeatedly since 2022. A buyer outside the Russian market should check reachability before designing around the price.

Related

russian-tts · yandex · russian-tts-evaluation · elevenlabs · text-to-speech · silero-tts