Multilingual voice form filling lets a visitor speak in their own language and fill a form written in another — automatic language detection first, then AI maps the spoken answers directly to your existing form fields. TypelessForm ships 25 interface languages, accepts spoken input beyond them, and normalises dates, phone numbers and addresses into the format each field expects.
This guide covers the whole pipeline — audio capture, speech recognition with language identification, slot filling, locale-aware normalisation, review — lists every interface language we ship, and compares the tools that fill the form you already have against the ones that replace it. Product claims below come from public vendor material and our own published documentation; where something is not published, we say so.
Disclosure: this guide is published by the TypelessForm team. TypelessForm is a site-side multilingual voice form filling widget. We describe alternative approaches as accurately as their public material allows.
What Is Multilingual Voice Form Filling?
Multilingual voice form filling means the visitor speaks naturally in their own language and AI extracts the entities — names, addresses, dates, quantities — and places them into the correct form fields, regardless of which language the form itself is written in. The defining capability is cross-language filling: spoken input in one language, form labels in another, correct result.
The base case of voice form filling is monolingual: an English speaker fills an English form. International websites face the harder case daily. A German customer lands on an English-language wholesale form; a Polish traveler books on an English hotel page; a Spanish-speaking applicant meets an English registration form. They can read the form — but composing answers, in writing, in a second language, on a phone keyboard, is where completions die.
Cross-language filling removes that step: the visitor speaks German, and the AI maps "Ich heiße Stefan Weber, Lindenstraße 8, 50674 Köln" into the name, street, postal code, and city fields of the English form. The visitor reviews the filled form and submits.
Why Are International Forms Harder to Complete?
International forms stack three frictions: typing in a non-native language, locale formats that differ from the visitor's habits (address order, postal codes, phone prefixes, date order), and mobile keyboards — where completion already runs 30% lower than desktop (Formisimo, 2023). Voice input removes the first and third friction; locale-aware normalisation handles the second.
Each friction compounds the others:
- Second-language composition. Reading a form in English is far easier than writing answers in it. Visitors who would speak fluently still hesitate to type, and hesitation on a 10-field form becomes abandonment — 68% of users already abandon forms overall (Baymard Institute, 2024).
- Locale mismatches. Address order differs by country (street-first vs city-first), postal codes vary in format and position, phone numbers need country prefixes, and date order (day/month vs month/day) is a classic silent error. A visitor's habits fight the form's expectations.
- Mobile keyboards in the wrong layout. An international visitor often types on a keyboard layout that lacks easy access to the diacritics or characters their answers need — names like "Müller" or "Gómez" get mangled or anglicized.
The safest pipeline, and the one TypelessForm implements, is transcribe-then-validate: speech becomes structured values first, then each value is normalised against the field's expected format, with a visible review step before submission. Voice should never bypass validation — it should feed it.
How Does Multilingual Voice Form Filling Work? ASR, Language Identification, Slot Filling and Locale Normalisation
Five stages: audio capture in the browser, automatic speech recognition with language identification, LLM slot filling that maps the transcript onto the form's detected fields, locale-aware normalisation of the extracted values, and a human review step before submit. TypelessForm ships all five behind one script tag; building it yourself means assembling them.
Each stage is a place where a build-your-own implementation either buys a vendor or writes code. Naming them separately is the only way to compare tools honestly, because most products in this category ship two or three of the five and leave the rest to you.
Stage 1 — Audio capture in the browser
One microphone button sits on the form, not on each field. The visitor presses it once and speaks a whole sentence — or several — covering as many fields as they like. This is the difference between one-shot filling and dictation: dictation asks the visitor to focus a field, speak, focus the next field, speak again. One-shot filling asks for one utterance and populates every field it can find. On a 12-field international checkout that is one interaction instead of twelve.
Stage 2 — Speech recognition and language identification (ASR + LID)
The audio goes to OpenAI Whisper, which transcribes it and identifies the spoken language in the same pass. There is no language selector and no per-language configuration on the site owner's side: the language is inferred from what the visitor actually says. This is what "automatic language detection" means in practice — the visitor never declares a language, and a site serving twelve markets installs the same script tag once.
Language identification also matters downstream. It is the signal that tells the normaliser in Stage 4 which conventions to read the spoken values under — whether "the fifth of March" arrived from a speaker whose locale writes dates day-first, and which country prefix an unprefixed phone number most likely belongs to.
Stage 3 — Structured entity extraction (slot filling)
The widget detects the form's fields from the DOM — names, labels, types, placeholder text, validation attributes — and hands that schema to GPT-4o together with the transcript. The model performs slot filling: it decides which fragment of the sentence belongs in which field. Nothing is matched positionally, so the visitor can answer in any order and skip anything they do not want to give.
Spoken (German), one press of the microphone:
"Ich heiße Stefan Weber, Lindenstraße 8, 50674 Köln,
Liefertermin der fünfte März, Telefon 0176 4433221."
Slot filling onto the English form's detected fields:
full_name -> Stefan Weber
street_address -> Lindenstraße 8
postal_code -> 50674
city -> Köln
delivery_date -> 2026-03-05 (from "der fünfte März")
phone -> +49 176 4433221 (from "0176 4433221")The last two lines are Stage 4, shown here so the hand-off is visible: slot filling decides where a value goes, normalisation decides what shape it arrives in.
Stage 4 — Locale-aware normalisation
Extracted values are rewritten into the format the field expects before they are placed. This is the stage most voice tools omit, and it gets its own section below because it is where multilingual input either becomes clean data or becomes a silent error.
Stage 5 — Review before submit, and what is never captured
The filled form is shown to the visitor, who can correct anything, and nothing is submitted automatically. Voice accelerates input; it never submits on the visitor's behalf. Alongside that, passwords, credit card numbers, CVV codes, and banking fields are excluded from voice capture by default, and no voice recordings are stored. On a form that collects personal data across borders, those two defaults matter more than any accuracy figure.
Assembled by hand, this pipeline is an ASR vendor (Whisper, Deepgram, Azure Speech, Google Cloud Speech-to-Text) plus your own DOM detection, your own extraction prompts and schema, your own per-country normalisers, and your own review UI — then maintained as forms change. TypelessForm is that pipeline pre-assembled behind one script tag. If you are weighing the build, our Web Speech API comparison and widget vs form-builder comparison cost it out.
Which Languages Does TypelessForm Support? All 25 Interface Languages, Listed
TypelessForm ships 25 interface languages — the widget's own buttons, prompts and error text — listed in full below. Spoken input is not limited to that set: language identification runs on Whisper, a multilingual speech model, so a visitor can speak a language the interface is not translated into and still fill the form correctly.
These are two different axes, and most published language counts blur them. The interface axis is what we ship and can enumerate; the spoken-input axis is what the speech model can hear.
| Language | ISO 639-1 | Language | ISO 639-1 |
|---|---|---|---|
| Bulgarian | bg | Norwegian | no |
| Chinese | zh | Polish | pl |
| Croatian | hr | Portuguese | pt |
| Czech | cs | Romanian | ro |
| Danish | da | Russian | ru |
| Dutch | nl | Serbian | sr |
| English | en | Slovak | sk |
| Finnish | fi | Spanish | es |
| French | fr | Swedish | sv |
| German | de | Ukrainian | uk |
| Greek | el | ||
| Hungarian | hu | ||
| Italian | it | ||
| Japanese | ja | ||
| Korean | ko | ||
Where this list comes from. Each language above is a locale bundle shipped inside the published widget package, so the count is checkable rather than a marketing figure. Anyone can verify it without our help:
npm pack typelessform-widget
tar -xzf typelessform-widget-*.tgz
ls package/dist | grep -E '^[a-z]{2}-'That returns one bundle per interface language — 25 of them in typelessform-widget 1.0.7. Two languages we are sometimes asked about, Arabic and Hindi, belong on the spoken-input axis rather than this one: Whisper recognises both, but the widget interface is not translated into either, so neither appears in the table above.
Why the language count is the wrong thing to compare
Published counts in this category — 25, 50+, 80+, 90+ — mostly measure the upstream speech model, not the product built on it. Almost every voice tool on the market routes audio through one of the same handful of multilingual ASR services, so a bigger number usually means the vendor is quoting their supplier's coverage rather than something they built. The count tells you what can be heard. It tells you nothing about what happens to the sentence afterwards: whether it lands in your existing fields, whether it crosses a language boundary, and whether the date ends up in the right order.
Cross-language filling: speak German, fill an English form
Cross-language filling is the case where the spoken language and the form's language differ. A German wholesale buyer opens an English order form, presses the microphone, and says the sentence in the Stage 3 example. The form is labelled "Full name", "Street address", "Postal code", "City", "Delivery date", "Phone" — and it fills correctly, because slot filling operates on meaning rather than on string matching between the transcript and the labels.
This is what separates speech-to-form from dictation. Dictation converts sound into text in the language it heard, and a German sentence typed into an English form is still a German sentence in the wrong boxes. Speech-to-form converts intent into structured, normalised values that satisfy the form's own rules. For an international site, that difference is the entire product.
Locale-Aware Formatting: Dates, Phone Numbers and Addresses
Recognising the words is the easy half. The value is converting what was spoken into the format each field expects — day-first or month-first dates, E.164 phone numbers with the right country prefix, postal codes and address order that differ by country. Normalisation runs before the review step, so the visitor confirms already-formatted values.
Three classes of value cause almost all of the damage, and each fails silently — the form validates, the visitor submits, and the error surfaces days later in fulfilment or support.
Dates
"The fifth of March" and "March fifth" are the same date spoken two ways, and 04/05/2026 is 4 May in Berlin and 5 April in New York. A wrong date passes every validation rule a date field has — it is a real date, correctly formatted, and simply not the one the visitor meant. Normalisation resolves the spoken form against the identified language and the field's expected pattern, then writes an unambiguous value, so a delivery date spoken in German does not arrive as a booking two months out.
Phone numbers
Visitors speak their number the way they dial it at home: a German says "null eins sieben sechs …" with a leading zero and no country code; a Briton says "oh seven double seven"; a Pole groups the digits in threes. Normalisation converts the spoken national form into E.164 with the correct prefix — 0176 4433221 becomes +49 176 4433221 — including the spoken conventions ("double seven" is two sevens, not the words) that raw transcription leaves as text.
Addresses
Address order is not universal. A US address runs street, city, state, ZIP; a Japanese address runs postal code, prefecture, city, then the finer subdivisions; a German address puts the house number after the street name. Postal codes differ in shape as well as position — 50674 in Germany, SW1A 1AA in the United Kingdom, 00-001 in Poland — so a single regex on the transcript finds the wrong fragment more often than not. Normalisation places each part into the field the form actually has, and diacritics survive the trip: Müller stays Müller and Gómez stays Gómez, rather than being flattened by a keyboard layout the visitor does not own.
One smaller case worth naming: decimal separators. A quantity or price spoken as "one point five" and one spoken as "eins Komma fünf" are the same number, and 1,5 in a field that expects 1.5 is another silent error.
All of this happens before the review step, which is the point. The visitor is not asked to check a raw transcript and mentally reformat it — they are shown finished, correctly shaped values and asked to confirm them.
Three Ways to Add Voice Input to an International Form
A site-side widget is installed once by the site owner and works for every visitor in 25 interface languages. A browser extension only helps the visitors who installed it themselves. The Web Speech API gives raw transcription in whatever languages the browser supports, with no cross-language filling and all parsing left to your team.
| Approach | Who installs it | Covers which visitors | Cross-language filling | Locale normalisation |
|---|---|---|---|---|
| Site-side widget (TypelessForm) | The site owner, once | Every visitor, every device | Yes — speak German, fill an English form | Built in, with review step |
| Browser extension | Each visitor, individually | Only visitors who installed it | Varies by extension | No — types raw text into fields |
| Web Speech API (build your own) | Your dev team builds it | Every visitor, behavior varies by browser | No — transcription only | Only what you build |
Site-side widget. One script tag added by the site owner puts a microphone on the form for every visitor. TypelessForm ships 25 interface languages with cross-language filling, runs all five pipeline stages above, and applies PII-safe defaults — passwords and card fields are excluded, no voice recordings are stored, GDPR-compliant. This is the only approach where the international experience is something the site owner actually ships, rather than hopes for.
Browser extensions. A category of user-side tools that add dictation to any text field. They genuinely help the individuals who install them — but for a site owner they are not a strategy: you cannot make your visitors install an extension, and the extension types raw transcribed text without knowing your form's locale rules. If your analytics show international abandonment, an extension on someone else's browser will not move it.
Web Speech API. The browser's built-in speech recognition is a real starting point for a custom build — it is free and needs no external service. Its limits are documented in our detailed comparison: it returns raw text for one field at a time, language support varies by browser, there is no cross-language filling, and the hard parts — parsing names, dates, phone numbers, and addresses across many languages — are all left to your team to build and maintain.
TypelessForm vs Other Multilingual Voice Tools
Most tools grouped under "multilingual voice" are voice agents or chatbots that run their own conversation, or builders that replace your form with theirs. A smaller set fills the form you already have. That distinction, plus locale-aware formatting, separates these products far more than the language count each one advertises.
| Tool | Languages published | Fills your existing form? | Cross-language filling | Locale-aware formatting | What it is |
|---|---|---|---|---|---|
| TypelessForm | 25 interface languages (listed above); spoken input beyond them | Yes — one-shot fill on the form already on your page | Published | Published — dates, phone numbers, addresses | Site-side voice widget for existing HTML forms |
| AnveVoice | "50+ Languages with automatic detection" | Yes — publishes agentic form completion | Not published | Not published | Voice AI agent that also performs actions in the page |
| Ask Bennett | 20 — ten in the automatic multilingual experience, ten as dedicated single-language setups | Not published | Not published | Not published | Multilingual website voice guide — answers visitor questions from approved site content |
| ChatLab | "Supports 80+ languages" | No — collects data in its own chat via "Custom Forms" | Not published | Not published | No-code AI support agent across chatbot, email, voice and WhatsApp |
| vForm | Not published | No — replaces your form with a voice form it builds | Not published | Not published | Voice-native form builder for lead qualification and discovery |
How these cells were filled. Every competitor claim above was read from the vendor's own live pages on 8 September 2026 — AnveVoice's multilingual detection page, Ask Bennett's multilingual page and homepage, ChatLab's homepage, and vForm's homepage — not from a summary or from memory. Where a capability is not documented we write "Not published" rather than "No": a vendor may ship something they have chosen not to write about, and we will not claim otherwise on their behalf. If any of these are out of date, tell us and we will correct the row.
Two of these deserve a fuller reading. AnveVoice is the closest comparison, and it publishes a larger language number than we do; what its multilingual page does not claim is either cross-language filling or locale-aware formatting, which is the axis this article is about. It also performs agentic actions inside the page, which we deliberately do not — a real capability difference, in its favour, depending on what you need. Our full side-by-side is in TypelessForm vs AnveVoice. vForm is a different product class rather than a weaker one: if you are willing to replace your form, a voice-native builder is a legitimate choice — see TypelessForm vs vForm AI.
For pricing, setup time and the wider widget field — including tools without a multilingual angle — see our comparison of voice form filling widgets.
How Do You Add Multilingual Voice Filling to Your Site?
Installation is one line of HTML on the page that holds the form — no backend changes, no rebuilding the form, no per-language configuration. The sensible rollout for an international site is to start with the form where your analytics show the highest non-domestic abandonment and run the free pilot (200 fills) against that baseline.
<script type="module"
src="https://cdn.jsdelivr.net/npm/typelessform-widget@latest/dist/typelessform.js">
</script>
<typeless-form api-key="YOUR_API_KEY"></typeless-form>The widget auto-detects the form's fields and adds the microphone; visitors speak and the language is identified automatically, with no language selection step and no per-market configuration. It works in React, Vue, Angular, WordPress, Shopify, and plain HTML. Pricing is published: free pilot at 200 fills, then $29, $99, and $199 per month, purchased via invoice.
A One-Line Decision Rule
Own the website and want every international visitor covered on the forms you already have? Install a site-side widget such as TypelessForm. Filling other people's forms as an individual? A browser extension serves you. Building a custom pipeline with a dev team? Start from the Web Speech API, knowing cross-language filling and locale normalisation are yours to build.
- Site owner, international audience, existing forms → TypelessForm — one script tag, 25 interface languages, free pilot of 200 fills.
- Site owner willing to replace the form itself → a voice-native form builder is a fair alternative; compare on data destination and locale handling, not language count.
- Individual user filling forms across the web → a browser extension, chosen for your own languages.
- Dev team building a custom voice pipeline → Web Speech API plus your own ASR, extraction and normalisation — see the build-vs-buy comparison first.
Related Reading
- What Is Voice Form Filling? — the base concepts behind speech-to-form.
- One-Shot Form Filling — why one utterance beats field-by-field dictation.
- Best Voice Form Filling Widgets — pricing, setup time and the wider field.
- TypelessForm vs Web Speech API — what building it yourself actually involves.
- Voice Input and Form Abandonment — the data on where voice input pays off, including multilingual users.
- One-Shot Voice Form Filling for E-Commerce — the same technology applied to store funnels.
Conclusion
International websites do not have a typing problem in general — they have a typing problem concentrated in visitors working outside their own language and locale. Multilingual voice form filling addresses that exact concentration: the visitor speaks naturally, language identification handles the rest, and the form fills in its own language and its own formats. The comparison that matters is not how many languages a tool can hear — almost everyone routes through the same speech models — but whether it fills the form you already have, crosses the language boundary, and gets the dates, phone numbers and addresses into the shape your systems expect. TypelessForm publishes all three, in 25 interface languages, with a free pilot of 200 fills.