Offline Hindi voice typing on a PC, with YazSes¶

What this page is: a how-to for pointing YazSes at Whisper's multilingual models. You hold a key, speak, release, and the text appears where your cursor is. Nothing is uploaded by default, there is no account, and there is no per-minute charge.

Most Hindi voice typing tools are websites or Android keyboards: you dictate into their text box and then copy the result out, and they stop working the moment your connection does. YazSes is a background program on your own computer, so it types into Word, a browser, a code editor, WhatsApp Web or a terminal directly, and it works on a train, in a village with patchy data, or on a machine that never touches the internet.

Which Indian languages Whisper can decode¶

The model shipped with YazSes is base.en — English-only. It cannot decode any of the languages below. You must switch to a multilingual checkpoint first (see the configuration section). Once you have, these are the language codes Whisper accepts; this project has measured none of them, and accuracy varies a great deal between them — larger models help disproportionately for lower-resource languages:

Accuracy is not the same across all of them — see honest limits below. Hindi, Bengali, Tamil and Urdu have more training data behind them than Assamese or Sanskrit.

Setting it up for Hindi¶

Two settings matter. The default model is English-only, so you must switch to a multilingual one:

[stt]
model = "small"        # NOT "small.en" — the .en models cannot do Hindi at all
language = "hi"

Then restart:

yazses restart

That is the whole configuration. The model downloads automatically the first time the daemon starts with it, and after that it never needs the network again. To fetch it as a separate, visible step (recommended behind a firewall or proxy), run yazses model download small first.

For another Indian language, change language to the code from the table above:

# Tamil
[stt]
model = "small"
language = "ta"

Which model size to choose¶

Indic languages benefit disproportionately from a larger model — more than English does. If Hindi output looks roughly right but keeps garbling word endings, the model size is usually the cause, not your microphone.

See performance tuning for the latency trade-off.

Hinglish — sentences that mix Hindi and English¶

Mixing English words into a Hindi sentence is normal Indian speech, and it is the case that trips up most dictation tools. For Hinglish today, use a multilingual model with auto-detection:

[stt]
model = "small"
language = ""          # auto-detect

Whisper then detects one language per burst rather than switching inside it, so mixed sentences come out best when the model is large enough to transcribe the English words in Hindi context.

[polyglot] is not the answer yet: it is planned, not wired. Its design routes a burst that mixes a declared pair ([polyglot] pair = "hi-en") to a separately trained code-switch adapter ([polyglot] adapter_path), but nothing yet detects where a burst switches language. Every burst is decoded by the [stt] model, with or without an adapter, so turning it on changes no transcript.

Will Devanagari actually type into my applications?¶

This is a fair question, because transcribing Hindi and typing Hindi are two different problems. YazSes has to synthesise keystrokes into whatever window has focus, and not every method can produce characters that are absent from your physical keyboard layout.

• Linux on X11 — works in the project's own end-to-end test. Devanagari types correctly through the xdotool backend, which maps each character to an X keysym rather than to a physical key: नमस्ते हिंदी dictated in, नमस्ते हिंदी received by the target application, byte for byte. That run is not yet a dated, versioned record in the validation matrix (#642).
• Any platform — the clipboard backend works by construction. It passes UTF-8 text straight through, so script is irrelevant to it:

[injection]
backend = "clipboard"

• Linux on Wayland — verify it yourself before relying on it. The default Wayland backend, ydotool, injects at the kernel input layer below the keyboard layout, so it can only press keys your active layout has. YazSes checks every non-ASCII character before typing: on a Latin keyboard layout there is no key for a Devanagari character, so YazSes pastes the whole burst from the clipboard instead ([injection] fallback_to_clipboard, on by default). That overwrites your clipboard and does nothing in most terminals. With a Devanagari layout active, the characters are typed. We have not confirmed Devanagari output end to end on Wayland; dictate one Hindi sentence into a text editor first.

If your text is transcribed but nothing lands in the application, that is an injection problem, not a language problem — troubleshooting covers it.

Honest limits¶

• Test it on your own voice before depending on it. Whisper's accuracy varies by language and by accent, and Indian-language performance is well behind its English performance. This page is not claiming parity.
• Smaller Indian languages are genuinely weaker. Assamese, Sindhi, Sanskrit and Nepali have far less training data than Hindi. A larger model helps but does not erase the gap.
• Multilingual models are slower than the English-only ones. On a modest CPU, medium in Hindi is meaningfully slower than base.en in English.
• Wayland + Devanagari is unverified, as described above. X11 and clipboard injection are the paths we have actually tested.

Related¶

• Multilingual dictation — code-switching, transliteration and translation in depth
• Personal vocabulary — names and terms the model keeps mishearing
• Install on Windows · Install on Linux · Install on macOS
• Performance tuning — model size vs latency