AI Dictation vs Diktuy: Russian-English Accuracy Test

AI Dictation produced the more faithful bilingual transcript in our Russian-English test. It scored 12.3% word error rate and preserved 26 of 27 English tokens. Diktuy scored 20.5% WER and preserved 16 of 27.

Diktuy did not fail outright. It completed the recording and returned readable Russian. Its weakness was more specific: it repeatedly converted the speaker's English words into Cyrillic or replaced them with Russian-looking equivalents.

Disclosure: AI Dictation is our product. Both scores use the same human reference and normalization method. This is one difficult recording, not a universal ranking of either app.

What did the comparison measure?

The source is the 70-second fashion interview used in our full bilingual dictation benchmark. The speaker uses Russian grammar while inserting English words such as "Mango colors," "capri pants," "on design," "jacket," "fit," "fabric," "price," and "spring and summer."

We compared each output with a 146-word human transcript and measured:

1. Word error rate, or WER. Lower is better.
2. English-token recall. How many of 27 English words remained in Latin script. Higher is better.
3. Completeness and fidelity. Whether the app reached the ending, preserved the proper name, and avoided translation or transliteration.

AI Dictation made 18 edits against the 146-word reference. Diktuy made 30. The wider difference appeared in English-token recall: AI Dictation lost one Latin-script token, while Diktuy lost eleven.

Full AI Dictation vs Diktuy scorecard

The comparison plan covers more than transcript accuracy. We include all twelve criteria below and mark unfinished controlled tests instead of replacing them with estimates.

Resident memory is a point-in-time process measurement, not a guaranteed requirement. Diktuy's range reflects multiple idle snapshots as its dashboard finished loading. We have not added active-recording memory until all products can be sampled during the same scripted run.

What did Diktuy get right?

Diktuy completed the clip and retained several of its most visible English phrases. It kept "Mango colors," the first "capri pants," "universal," both instances of "amazing," "look couture," "couture prices," one instance of "details," "affordable," and "spring and summer."

Here is the captured output:

Что модно в этом сезоне, на ваш взгляд и на взгляд, Саймон Ченек? Mango colors. Шелк очень модный. Очень модные пиджаки с какими-то деталями. Брюки модно. Capri pants модно. И это все он дизайн. Джинсы он очень красиво и делает. Можно надеть капри джинсы, красивую блузку под низ, накинуть джакет. Можно так на работу тоже выйти. Можно так пойти погулять потом. Так что он universal. И модно все. Все красивое и фит его amazing. Amazing на всех. У него look couture без этих couture prices. Его костюмы, смотря какой фабрик, смотря откуда приходит фабрик, смотря сколько details в этом пиджаке. А может на прайс от 225, самый дорогой 380. Так оно очень-очень affordable. Да. Ну что ж, замечательно. Ну и в заключении чисто традиционный вопрос. Ваши пожелания нашим милым обаятельным телезрительницам. Я вам желаю хорошего spring and summer, чтобы вы приходили, красиво одевались и мы вам можем помочь.

As Russian prose, much of that is understandable. As a record of bilingual speech, it needs repair.

Which English words did Diktuy lose?

Diktuy flattened language boundaries in several ways:

The proper name also changed. The human reference identifies the designer as Simon Chang, inflected in the Russian sentence as Simon Чена. Diktuy returned "Саймон Ченек," changing both the script and the surname ending.

Some differences are less important. Diktuy changed "одеть" to the grammatically preferred "надеть" and added punctuation. Those edits may improve written Russian, but they also show that the output is not strictly verbatim. In this benchmark, fidelity matters more than polish.

Why is English-token recall separate from WER?

Diktuy's 20.5% WER is better than Wispr Flow's 27.4% score on the same recording, yet Diktuy preserved fewer English tokens: 59.3% versus 63.0%.

That is possible because WER treats every word edit equally. Replacing one Russian word and transliterating one English word each cost an edit, even though the second error may matter more to someone who intentionally code-switches. English-token recall isolates that behavior.

A transcript can therefore look fluent, remain close to the Russian sentence structure, and still erase the speaker's choice of language. Bilingual dictation needs both measures.

How did AI Dictation handle the same clip?

AI Dictation preserved 26 of the 27 English tokens. It kept "Mango colors," both instances of "details," "capri pants," "on design," jacket, universal, fit, both instances of amazing, "look couture," "couture prices," both instances of fabric, price, very, affordable, and "spring and summer."

Its missing English token was Simon, which appeared in Cyrillic as "Саймон." The transcript still contained Russian wording differences, so the 12.3% WER should not be read as perfect accuracy. It was simply much more faithful to the language switching than Diktuy's output.

AI Dictation's cloud mode lets users enable both languages and tells its cleanup stage to preserve code-switching. That does not guarantee every borrowed word will survive, but it reduces the pressure to normalize a mixed sentence toward one script.

Is Diktuy a native Mac app?

Diktuy 2.6.2 is a signed and notarized Apple silicon macOS app, but its interface is not built in SwiftUI or AppKit. The main executable is a PyInstaller package that embeds Python 3.11 and PyQt6. Qt draws the cross-platform desktop interface through its macOS platform layer.

The difference is visible in the bundle size. Diktuy's installer was 397 MB and the installed app occupied about 1.1 GB. AI Dictation occupied 32 MB. The largest Diktuy component was a 762 MB embedded "Diktuy Recording Helper" app containing another Python, Qt, PyTorch, ONNX, Sherpa, NumPy, and SciPy stack. Diktuy's own runtime log for version 2.6.2 said macOS was using single-process mode and skipping that helper, so much of the duplicated helper bundle was not loaded during our idle measurement.

The main process still includes substantial local audio tooling: Silero voice-activity detection, ONNX Runtime, Sherpa ONNX, FFmpeg, PortAudio, NumPy, and SciPy. During a test recording, its log showed local 16 kHz microphone capture, Silero VAD initialization, and a warmed API connection. Diktuy's public site says voice dictation uses Whisper Large-v3-turbo, while uploaded file and video transcription can run locally without limits. In other words, the package contains local preprocessing and offline file-transcription components, but normal voice dictation is an account-backed cloud workflow.

The dashboard offers history, file transcription, a dictionary, text replacement, agents, and settings. That breadth helps explain both the larger interface and the heavier package. It is a cross-platform Qt application with native macOS hooks, not a native SwiftUI/AppKit app.

Should Russian-English users choose AI Dictation or Diktuy?

AI Dictation is the stronger choice for the speech pattern tested here. It made 12 fewer word edits and preserved ten more English tokens in Latin script.

Diktuy remained usable if the goal was understandable Russian text and the user was willing to repair transliterated terms. It was less suitable when English spellings, product names, fashion terms, or exact bilingual wording needed to survive unchanged.

This result should stay in proportion. It covers one speaker, one recording, and one Russian-English style. Diktuy may perform differently on monolingual Russian, another accent, or a different device. The test is decisive only for this job: preserving dense code-switching inside otherwise Russian speech.

Diktuy's pricing is unusually low in its home market. The free plan includes 30 dictation minutes per month. Pro lists 300 monthly minutes at 299 ₽ month-to-month or 239 ₽ per month billed annually. Max lists unlimited dictation at 599 ₽ month-to-month or 479 ₽ per month billed annually. AI Dictation Pro is $8.49 per month for unlimited dictation, so a direct price comparison depends on billing currency and which non-dictation features matter.

For more results, read the full multilingual dictation benchmark, the broader best voice-to-text software guide, AI Dictation vs Soniox App, and AI Dictation vs Oravo. To test your own vocabulary, download AI Dictation, enable the languages you use, and speak normally.