Dictating in more than one language¶

Short answer: the underlying Whisper model is multilingual, so if you supply a multilingual checkpoint yourself, YazSes will pass your chosen language through to it. On top of that YazSes adds some conveniences bilingual speakers ask for — translating as you speak and transliterating Persian into its script. Switching language per application and routing sentences that mix two languages mid-thought are designed but do not work yet; the sections below say exactly how far each one goes. The recognition quality in any of those languages is Whisper's, and we have not measured it.

All of it runs on your machine. This matters for languages where the cloud alternatives are weakest or absent entirely.

Dictating in one non-English language¶

Set the language in your config and use a multilingual model — note that the .en models (base.en, small.en) are English-only by design, so switch to the multilingual variant:

[stt]
model = "small"        # not "small.en" — the .en models are English-only
language = "de"        # or fr, es, it, fa, ar, hi, zh, …

Then yazses restart. Accuracy varies substantially by language and model size; larger models help disproportionately for lower-resource languages.

For Mandarin there is a one-command route that writes every key and fetches the model: yazses language set zh-CN (or zh-TW); add --dry-run to preview it. See Chinese voice typing. yazses language list shows the profiles it supports.

German, French, Spanish¶

These three are the most common request, so here they are as complete configs. Drop the block that matches your language into config.toml and restart:

# German
[stt]
model = "small"
language = "de"

# French
[stt]
model = "small"
language = "fr"

# Spanish
[stt]
model = "small"
language = "es"

Downloading the model¶

faster-whisper fetches small (or whichever multilingual model you configured) automatically the first time the daemon starts with it set, the same way it fetches .en models, and caches it in the Hugging Face cache. To fetch it as a separate, visible step (recommended behind a firewall or proxy), run yazses model download small first. Multilingual weights are somewhat larger than their .en counterparts at the same size tier, so expect a longer first-run download and a few hundred MB more disk use. Confirm it landed with:

yazses doctor

which reports the configured model as cached, local, or not yet downloaded. If the download stalls or fails partway (a slow or interrupted connection is the usual cause), delete the partial cache entry and restart the daemon to retry from a clean state — yazses doctor output tells you which model is misbehaving. The cache lives under ~/.cache/huggingface/hub (or $HF_HOME/hub if you set that), one models--… directory per model, so removing just the offending one leaves your other models intact.

Letting YazSes detect the language¶

If you switch languages often and don't want to edit config each time, leave the language empty and Whisper will detect it per utterance:

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

This costs an extra detection pass on every burst and can occasionally guess wrong on very short utterances, so prefer an explicit code when you know what you'll be speaking.

Sentences that mix two languages¶

This is the case ordinary dictation handles worst. Bilingual speakers routinely switch mid-sentence — a Persian speaker dropping in English technical terms, a Spanish speaker using English product names — and a single-language recogniser mangles whichever language it was not set to.

What works today is a multilingual model with auto-detection, which picks one language per burst:

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

[polyglot] is planned — not yet wired, so yazses features enable polyglot refuses it. The design (ADR-v2-008) routes a burst that mixes the two languages of a declared pair (pair = "fa-en") to a code-switch adapter you train separately ([polyglot] adapter_path). What is missing is the input that decision needs: nothing yet tells the daemon which parts of a burst are in which language, so every burst is decoded by the [stt] model, with or without an adapter. If an older config of yours turns it on, the daemon log and yazses doctor say so; yazses features disable polyglot clears it.

Per-application language switching¶

If you write in one language in your email client and another in your editor, switching by hand every time is friction that makes dictation not worth using. Planned — not yet wired, so features enable refuses it for now:

yazses features enable langroute

Language then follows the application you are dictating into.

Translating as you speak¶

Two distinct capabilities, often confused:

For an existing recording rather than live dictation, the same translation path is one flag:

yazses transcribe interview.fr.m4a --language fr          # French in, French out
yazses transcribe interview.fr.m4a --language translate   # any language → English

--language takes any Whisper code, and it now reaches the decoder: until v2.30 only translate did anything, so --language fr set nothing and the file was transcribed as English. A non-English code needs a multilingual model — drop the .en suffix from [stt] model (or pass --model small). YazSes refuses with exit code 2 and names the model to use, instead of handing you fluent English for non-English audio.

Non-Latin scripts¶

Two different problems here, and they need different tools:

• Transliteration (yazses features enable translit) — dictate Persian in romanized Finglish and get Persian script, by a letter-by-letter table lookup. Finglish is currently the only scheme.
• Diacritics (yazses features enable diacritize) — restores accents dropped from common loanwords (cafe → café, naive → naïve) from a fixed lexicon. It does not vocalise Arabic or Hebrew or restore Vietnamese tones.

There is also SafeGlyph (yazses features enable safeglyph), which guards against visually confusable characters — relevant when mixing scripts in identifiers or URLs.

Getting a non-Latin script to actually type¶

Transcribing a script and injecting it are separate problems, and they can fail independently. Text can be recognised perfectly and still not reach your application, because the injection backend has to synthesise characters that are not on your physical keyboard layout:

• X11 (xdotool) maps each character to an X keysym, so it handles non-Latin scripts. Verified end-to-end with Devanagari.
• Clipboard ([injection] backend = "clipboard") passes UTF-8 straight through — script-independent by construction, and the reliable fallback.
• Wayland (ydotool) injects 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 layout there is no key for a Devanagari or Arabic character, so it pastes the whole burst from the clipboard instead ([injection] fallback_to_clipboard, on by default), which overwrites your clipboard and does nothing in most terminals. Test one sentence before relying on it.

If the transcript is right but nothing appears in the target application, change the backend before changing the model.

Pronunciation feedback¶

If you are dictating in a language you are still learning, there is an optional capability that gives you feedback on pronunciation rather than silently transcribing something you did not intend. Planned — not yet wired, so features enable refuses it for now:

yazses features enable pronunciation

Smoke-testing your language, offline¶

Nothing above is a measurement. The presets are derived from model size and Whisper's published behaviour, not from anyone testing your language — and per-language quality varies far more than size alone predicts. There is no substitute for hearing your own voice come back.

Two ways, and the second needs no microphone once you have a clip:

# Live: hold your hotkey, say a sentence you can check.
yazses restart && yazses status

# From a file — reproducible, and shareable without re-recording.
yazses transcribe path/to/clip.wav

Making your own fixture¶

YazSes deliberately ships no recorded speech fixtures for non-English languages. A voice recording is personal data; licensing on public speech corpora varies clip by clip; and a fixture nobody may legally redistribute is worse than none, because it invites someone to add one that cannot stay. Record your own — it is a minute of work and it stays yours:

# 10 s of 16 kHz mono, the format the engine wants
arecord -f S16_LE -r 16000 -c 1 -d 10 my-clip.wav     # Linux
# macOS: QuickTime → export WAV.  Windows: Voice Recorder, then convert.

yazses transcribe my-clip.wav

Keep the clip out of the repository. To contribute a result rather than audio, use the table below.

Report what you found¶

Open an issue titled "Language report: " (https://github.com/MSKazemi/yazses/issues/new) with one row per language. This is the only way the project learns which languages are genuinely usable, because no one here speaks them all.