Voice notes for website inquiries

Visitors can describe a website problem by voice, review the transcript, and send the message from our own site.

Case type
Website feature
Business
Taranov Digital’s own website
Focus
Contact workflows, Voice transcription, Static websites

The task

Build a contact option for visitors who explain their situation more easily aloud. Turn the recording into a readable draft, with a separate review and send step, while keeping the marketing site fully static.

What the project needed to do

  • Let visitors describe their website problem in their own words, by voice or text.
  • Give visitors an editable message and control over the contact details before sending.
  • Keep dynamic processing separate from the static marketing website.

What we built

Record in the browser

Add English voice recording for up to three minutes, with text entry available if recording is unsupported or microphone access is blocked.

Review before sending

Upload the recording for transcription, return the transcript to the message field, and leave the draft editable. Sending remains a separate action.

Connect a static site to a focused service

Use a Cloudflare Worker for audio processing, private storage, and message submission. The Astro site keeps a small browser script for the form and recording controls.

Implementation walkthrough

Business type
Own product — the widget on this site
Location
The Woodlands, Texas
Started
2026
Last updated
July 19, 2026

Stack Cloudflare Workers, Cloudflare R2, Cloudflare Email Routing, Workers AI, Groq, Whisper large v3, Astro

Use the voice and text contact control in this page’s footer to try the feature. Record a message, review the transcript and any email or website extracted from it, then choose whether to send.

The site itself is a fully static Astro build with no server. Everything here runs in a single Cloudflare Worker on a separate subdomain, with Groq doing the speech and extraction work.

Why a voice note at all

The people this site is written for — owners of established service companies — describe their problem far better out loud than in a form field. Ask them to type and you get “need a new website.” Ask them to talk for ninety seconds and you get the actual situation: what the current site does wrong, which services matter, what happened with the last agency.

A text box collects the short version. The question was whether the longer version could be collected without asking anyone to install, sign up, or schedule a call.

An audio file is difficult to skim or search. Transcribing the recording before delivery makes the inquiry readable in an inbox or on a phone.

How it works

From spoken words to an email
  1. The visitor records in the browser

    No install, no permissions beyond the microphone, up to three minutes.

  2. The audio uploads to a private bucket

    Stored under an unguessable name, tagged with the page it came from.

  3. It is transcribed while the visitor waits

    Deliberately synchronous — the transcript has to come back, because it is the thing the visitor is about to edit.

  4. Contact details are pulled out of the speech

    A spoken email or website is normalized and, critically, checked against the transcript before it is trusted.

  5. The form fills itself in

    Transcript in the message box, email and website in their fields. Everything is editable — nothing has been sent yet.

  6. The visitor edits and submits

    This is the moment the email is sent, not the moment they finished speaking.

  7. The note arrives as readable text

    Plus a private link to hear the original in the sender's own voice, and a reply-to address so hitting reply just works.

Review the draft before sending

Transcription happens on upload, before the visitor submits anything. That costs a wait of a few seconds — but it means the transcript is not a byproduct filed away somewhere, it is the draft the visitor then edits. Speech-to-text gets names and company names wrong; letting someone fix their own name before it reaches me is worth more than saving them the wait.

An abandoned recording is not emailed. The visitor must submit the form to send the message.

Readable email with an audio link

The note arrives as plain text: the message, the fields, the page it came from, the country. Underneath is one line with a link to hear the original recording, valid for thirty days.

The reply-to header is set to the visitor’s address when one was given or spoken, so replying works without copying anything out of the body. Tone, hesitation, and emphasis stay available in the recording for the cases where they matter — but reading comes first.

Architecture

The marketing site remains a static build. Recording, transcription, storage and message delivery run in a separate service, so the website itself needs no server runtime.

Technical architecture
  1. Static Astro site on Cloudflare Pages

    No server-side rendering, no adapter, no runtime on the site itself.

  2. One Worker on a separate subdomain

    Three endpoints: upload-and-transcribe, submit-note, and playback. All form, voice, and email logic lives here.

  3. R2 for the recordings

    Private bucket. Objects carry the originating page and country as metadata.

  4. Groq for transcription and extraction, Workers AI as failover

    Whisper large v3 and gpt-oss-120b on Groq; a Workers AI binding takes over if Groq is unreachable, so an outage degrades the feature instead of breaking it.

  5. Cloudflare Email Routing to deliver

    Multiple destinations attempted in order, so one misconfigured mailbox is not a lost lead.

The browser talks to a different origin than the page it is on, which makes CORS part of the design rather than an afterthought. Requests from the site’s own origins are allowed; requests carrying no origin at all — a click on a playback link inside an email client, for instance — are also allowed, because that is a legitimate path that no browser will label.

Documented in depth

Browser support and fallback behavior

Without JavaScript, the contact form is an ordinary HTML form that posts to the Worker and redirects back to a URL fragment, where a CSS rule reveals the confirmation banner. No script runs at any point. The voice recorder is not available in this mode, but sending a message is.

On iPhone, Safari records in a different audio container than Chrome does. Both are accepted, along with two more formats, because a visitor should never be told their phone is unsupported.

When transcription fails entirely, the recording still uploads, the note still sends, and the audio link still works. The transcript is simply empty. Every AI call in this system is allowed to fail without taking the message with it — the point is to deliver a lead, not to demonstrate a model.

Upload and submission safeguards

A hidden field is present in the form that no human ever sees. Bots fill in every input they find, so a submission with that field populated gets a cheerful success response and is silently discarded — no error to learn from, no signal to adapt to.

Uploads are bounded at both ends: too small to be speech is rejected, too large is rejected, and only four audio types are accepted. Field lengths are capped. A submission with no way to reply at all is refused.

Playback links carry no signature and no expiry parameter. The link is the credential: a random name long enough to be unguessable, with age checked against the file’s own upload time on the server. There is nothing in the URL to tamper with.

Automated rate limiting is not built. Volume is currently low enough that the honeypot and size limits are doing the work, and adding a limiter before there is abuse to shape it against would be guesswork.

Implemented features

This system runs in production on this very site — the widget on this page is it. The one item not built is automated rate limiting, and it is marked as such.

Component status

9 live · 1 planned

Component Status Note
Browser recording, up to three minutes Live Four audio formats including iOS Safari
Synchronous transcription Live —
Transcription provider failover Live Second provider on a different platform
Contact extraction verified against the transcript Live —
Editable prefill before sending Live —
Private playback link, thirty-day life Live —
Email delivery with reply-to and destination failover Live —
No-JavaScript form fallback Live CSS-only confirmation, no script
Honeypot and size validation Live —
Automated rate limiting Planned — not built Not built — no abuse pattern to design against yet
Status at the last update to this page.

Implementation problems and fixes

Challenge and lesson

The model invented contact details

Initial assumption
Pulling an email address out of a transcript is a trivial extraction task, so the cheapest, fastest model is the obvious choice.
Hidden complexity
Given a transcript where someone named their website but never said an email, the small model returned one anyway — info@ or contact@ at that domain. It was not a hallucination in the obvious sense: it was a reasonable guess, correctly formatted, and completely wrong. Worse, it was invisible. A fabricated address looks exactly like a real one in a form field, and the visitor, seeing their own website spelled correctly beside it, has every reason to trust the form and submit.
Architectural response
Two changes. A larger model that follows a strict output schema replaced the small one. More importantly, a rule was added that no prompt can override: an extracted value is discarded unless it literally appears in the transcript, after normalizing spoken forms like 'at' and 'dot'. The check is mechanical, it runs on every result, and a dropped value is logged.
Lesson

Do not instruct a model not to invent things and consider the problem solved. Where an output can be checked against its own source, check it in code. Prompt wording is a preference; a verification step is a guarantee.

Challenge and lesson

Playback links were too long

Initial assumption
A private file needs a signed URL: a key, an expiry timestamp, and an HMAC proving the other two were not altered.
Hidden complexity
That produced links of roughly a hundred and eighty characters, two thirds of which was a signature. These links go into an email that a human reads. They wrapped across lines, looked like phishing, and carried an expiry the recipient could see but not act on — while the actual expiry logic still had to be verified server-side anyway.
Architectural response
The file's name became the credential: sixteen random bytes, which is unguessable, and nothing else. Expiry is derived from the object's own upload timestamp, which the storage layer already records. No key, no signature, no query string — a link short enough to read.
Lesson

Before signing a URL, ask what the signature is actually protecting. If the only secret is 'which file', a sufficiently random filename is that secret, and the server can answer every other question from data it already holds.

Challenge and lesson

Server errors looked like CORS failures

Initial assumption
Cross-origin headers are added to responses, and errors are handled where they occur, so the two concerns do not interact.
Hidden complexity
When an unexpected exception escaped the handler, the runtime returned its own bare 500 — which never passed through the code that attaches cross-origin headers. The browser saw a response it was not allowed to read, and reported a CORS failure. So every genuine backend error presented in DevTools as a configuration problem, sending debugging in precisely the wrong direction.
Architectural response
The entire router is wrapped so that any thrown error becomes a normal JSON error response that still travels through the cross-origin layer. The browser now receives a readable error saying what actually went wrong.
Lesson

On a cross-origin API, the error path needs the same headers as the success path. An error the browser cannot read is worse than no error at all, because it actively misdirects.

Outcomes

  • A visitor can describe their situation in the way that suits them, without typing it.
  • Notes arrive readable and searchable, with the original audio one click away when tone matters.
  • Replying works directly from the email, with no copying.
  • The marketing site stayed fully static, with all dynamic behavior in one small, separately deployed service.
  • The message remains usable if transcription or extraction fails.

Where this pattern fits

Anywhere the visitor knows more than the form is able to ask, and typing it out is the thing stopping them:

Here Elsewhere
“Describe the problem with your website” Intake for legal, medical, insurance, or repair work
Spoken email and website Any detail people say faster than they type
Transcript emailed to one person Transcript into a CRM, help desk, or shared inbox
Thirty-day audio link Retention that matches your record-keeping rules

The same approach can support other subjects: return an editable transcript before submission, verify extracted values against the source and allow the message to continue if an AI step fails.

Alex Taranov
Who built this

Alex Taranov mapped the workflow, designed the data model and architecture, and oversaw implementation. AI-assisted tooling supported the coding; the product decisions, review, testing, and final responsibility stayed with Taranov Digital.

About Alex

What the case demonstrates

The result is a voice-or-text contact form with a transcript the visitor can edit before submitting. The walkthrough above explains the recording, storage, and delivery design.

Discuss a workflow

Start with a focused module connected to your existing CRM, order system or scheduling tool. Tell Alex which workflow you want to improve.

Tell Alex about your website.

Share your website and what you would like to improve. Alex will reply with a short video review and two or three priorities.

Text Alex(281) 946-0946Email Alex[email protected]
Office26700 Kuykendahl RdThe Woodlands, TX 77375
Privacy policy

Request a Website Review