Forem: Digiwares

How Google De-indexed My Entire Site - and What I Did to Recover

Digiwares — Wed, 08 Apr 2026 12:39:59 +0000

Last November, Google quietly dropped almost every page of my site from its index. I went from having 200+ pages crawled to 1 page indexed — the homepage. It took me weeks to figure out what happened, and months of work to fix it.

This is the full story.

The Product

VoiceToTextOnline.com is a browser-based voice-to-text tool. It uses the Web Speech API — no server processing, no audio stored, nothing uploaded. You open the page, click a button, and speak. The browser does all the recognition locally.

It also supports file transcription (upload an MP3/WAV/MP4, get a transcript with speaker labels and SRT export) and text-to-speech via Google Cloud neural voices.

I launched it in early 2025 as a solo founder. By mid-2025 it was getting 150-200 daily visitors mostly from Bing and AI referrers like ChatGPT and Doubao (ByteDance's AI assistant). One paying Pro subscriber.

Then the GSC numbers started looking strange.

What Happened

I had made three mistakes simultaneously, and they compounded each other.

Mistake 1: www vs non-www split

At some point my deployment config started serving the site on both www.voicetotextonline.com and voicetotextonline.com without a proper canonical redirect. Google was seeing duplicate versions of every page and splitting crawl budget between them.

Mistake 2: Middleware blocking Googlebot

I had Next.js middleware handling auth redirects. The middleware was checking session state on every request — including requests from Googlebot. In certain conditions it was returning redirect responses to the crawler instead of the actual page content.

// The problematic pattern — middleware firing on all routes
export function middleware(request: NextRequest) {
  // This was intercepting Googlebot requests
  const session = getSession(request)
  if (!session) {
    return NextResponse.redirect('/auth')
  }
}

Googlebot would hit a page, get redirected, follow the redirect, potentially get redirected again. Some pages were returning 308 chains that never resolved cleanly.

Mistake 3: Sitemap pollution

My sitemap.ts was including URLs from non-Phase-1 pages that I had intentionally set to noindex via HTTP headers. So the sitemap was advertising pages that returned X-Robots-Tag: noindex — a contradiction that confused Googlebot's crawl prioritisation.

The combination of all three meant Google had spent months crawling a confusing, inconsistent site — duplicate URLs, broken redirects, noindex signals fighting sitemap inclusions. Its response was to stop trusting the site almost entirely.

Result: 220 pages crawled, 1 indexed.

The Fix (Technical)

I combined all three fixes into a single commit in early March 2026.

1. Force non-www canonical redirect

// middleware.ts — added before any other logic
const host = request.headers.get('host') || ''
if (host.startsWith('www.')) {
  const url = request.nextUrl.clone()
  url.host = host.replace('www.', '')
  return NextResponse.redirect(url, 301)
}

2. Exclude Googlebot from auth middleware

const userAgent = request.headers.get('user-agent') || ''
const isBot = /googlebot|bingbot|slurp|duckduckbot/i.test(userAgent)

if (isBot) {
  return NextResponse.next() // Let crawlers through unconditionally
}

3. Rebuild the sitemap to only include indexable pages

Removed all pages with noindex headers from sitemap.ts. The sitemap now only advertises pages that return 200 with no noindex signal.

I submitted a GSC validation request on March 17. It failed on March 21. The crawled-but-not-indexed count was still at 91 pages.

This is where the real problem became clear.

The Real Problem: Thin Content at Scale

Even with the technical issues fixed, Google wasn't indexing the language pages. And looking at them honestly, I understood why.

I had 43 language-specific pages — one for each language the tool supports. Hindi voice to text, Arabic voice to text, Japanese voice to text, and so on. Each page was built from the same template:

Hero with the language name
4-step how-to
Voice commands table
"Why Choose Us" 4-card section (identical on every page)
FAQ
CTA

The language name was swapped in. The content was largely identical. To Google's quality systems, this looked like 43 copies of the same page — thin content at scale.

The site:voicetotextonline.com search confirmed it. Only the homepage appeared.

The Content Fix: 43 Unique Pages

I spent several weeks rewriting all 43 language pages from scratch. The rule was simple: each page had to say something true and specific about that language that no other page on the site said.

This required actual research into the linguistics, diaspora communities, and cultural context of each language. A few examples of what "genuinely unique" ended up meaning:

Lithuanian — Lithuanian is considered the oldest surviving Indo-European language, more archaic than Latin or Sanskrit. Real Lithuanian-Sanskrit cognates exist and are fascinating: avis (sheep) in Lithuanian vs áviḥ in Sanskrit, sūnus (son) vs sūnúḥ. The page leads with this and shows a comparison table. The UK/Ireland diaspora section covers the 200,000+ Lithuanians who moved after EU accession in 2004 and face keyboard incompatibility with the 9 special Lithuanian characters.

Slovenian — One of the only living languages with a grammatical dual number — a distinct form for exactly two of something. The three-card visual showing ednina (singular) / dvojina (dual) / množina (plural) with real examples is content no competitor page has.

Catalan — 10 million speakers but no EU official status, despite having more native speakers than Irish, Maltese, and Luxembourgish combined. The comparison table makes this concrete. The l·l character (ela geminada) exists only in Catalan and is impossible to type on any standard keyboard — the page explains this and positions voice typing as the solution.

Bulgarian — The Cyrillic alphabet was created in Bulgaria in the 9th century. Every other Cyrillic language page can claim the script as their writing system. Only Bulgarian can claim they gave it to the world. The three-card section (9th century / 250M+ users / 50+ languages) is genuine and searchable.

Afrikaans — The youngest natural language in the world (standardised 1925) and one of the most accurately recognised by Web Speech API, because it has no special characters and writes almost exactly as it sounds. The comparison table shows why Afrikaans outperforms German, Danish, and Dutch for voice recognition accuracy.

After 43 rewrites, every page has:

A unique lead angle based on genuine linguistics
A named culturally-specific section (diaspora keyboards, minority status, historical alphabet events)
A sample output block showing real text in the target script
5 language-specific FAQs

The Internal Linking Gap

While doing this I discovered another issue: every language page's breadcrumb pointed to /#languages — an anchor that didn't exist on the homepage. 43 pages with broken breadcrumb links.

The fix was building a proper /languages index page linking to all 43 pages, grouped by region, with flag, native script, and speaker count for each. This also gave Google a clear crawl path: homepage → /languages → all 43 language pages.

Current Status

10 rewritten language pages submitted to GSC via URL Inspection
The /languages index page deployed and submitted
GSC still shows 1 indexed page (the homepage)
91 pages in "Crawled - currently not indexed" with failed validation
124 pages "Discovered - currently not indexed" with passed validation

The 124 "Discovered" pages are the hopeful signal — Google has them queued. The 91 failed pages are the harder problem: Google formed its quality judgment when the content was thin, and that judgment is sticky even after the content improves.

The honest timeline for recovery is 3-6 months from the content fix. The technical fixes were necessary but not sufficient. The content rewrites were necessary but not sufficient. What's likely needed next is external backlinks — at this domain trust level, Google needs signals from other sites before it will fully re-evaluate.

Lessons

1. www/non-www is not a minor detail. It's a canonical URL decision that affects every page on your site. Pick one and enforce it with a 301 at the infrastructure level, not just in meta tags.

2. Middleware that touches all routes will touch Googlebot. Any redirect logic in Next.js middleware needs a bot exemption or it will silently confuse crawlers.

3. Sitemaps and robots signals must be consistent. If a page is in your sitemap, it should return 200 with no noindex. If it's noindexed, remove it from the sitemap.

4. Template pages at scale look like duplicate content. 43 pages with the same structure and swapped keywords is not 43 unique pages. Each page needs to say something the others don't.

5. Google's quality judgment is sticky. Fixing the content doesn't immediately fix the index. Google cached its assessment when the pages were bad. Re-earning trust takes time and likely external validation.

If you're building a multi-language tool and want to avoid this — hope the technical details here save you some pain.

The tool is at voicetotextonline.com if you want to see what the rewritten pages look like. The Lithuanian and Slovenian pages are probably the most interesting from a linguistics angle.

Solo founder, building in public from Bangkok. This is month 13 with zero revenue. The goal is $5K MRR. Current status: 1 paying subscriber.

I Built a Free Voice-to-Text Tool That Supports 55+ Languages published: true

Digiwares — Tue, 03 Feb 2026 03:54:09 +0000

We developed VoiceToTextOnline.com in a period of 3 months and now on Month 4 as I type this post. On month 4, we are getting 200 visitors a day and bounce rate of 50 to 52%.

It was a lot of effort working on the SEO to drive traffic. Now my No.1 traffic source is Bing and we rank No.3 on there but ranking on Google is still a far away story! I even setup 2 satellite websites to drive traffic last week- speechtotext.online and texttospeech.site. They have been able to get around 10 to 20 unique visitors a day. In the last 3 months, we have a solid traffic base. The only worry point is we are still not even able to make it to the second or third page of Google SERP.

Based on user request we also setup a new tool on the website - text to speech / TTS though our site says voice to text online dot com.

I want to ask the fellow developers - what strategy do you use to rank on Google?

Introducing Voice to Text Online - a powerful speech recognition tool that works right in your browser.

🌍 55+ LANGUAGES SUPPORTED
Hindi • Spanish • Arabic • French • German • Italian • Portuguese • Chinese • Japanese • Korean • Russian • Turkish • Tamil • Telugu • Bengali • and 40+ more!

✨ WHY PROFESSIONALS LOVE IT:

→ No account needed — just open and start talking
→ Real-time transcription — see your words appear instantly
→ Works on any device — phone, tablet, or computer
→ No downloads required — 100% browser-based
→ Privacy-first — your voice data stays on your device

💼 PERFECT FOR:
📝 Taking meeting notes
✍️ Drafting emails and documents
🎙️ Transcribing interviews
📚 Creating content faster
🌐 Multilingual teams

The best part? It's completely free. No hidden fees, no premium walls for basic features.

👉 Try it now: voicetotextonline.com

I Built a Group Chat That Draws Itself Into a Graph - Here's How

Digiwares — Mon, 26 Jan 2026 09:38:35 +0000

I Built a Group Chat That Draws Itself Into a Graph - Here's How

Ever been in a group chat where great ideas just... disappear? You're brainstorming with your team, someone drops a brilliant insight, and 30 messages later it's buried forever.

I got tired of scrolling through walls of text trying to find "that thing someone said earlier," so I built Qonvo — a group chat that visualizes conversations as an interactive graph in real-time.

The Problem

Traditional chat is linear. But conversations aren't.

We jump between topics, reference earlier points, and connect ideas across threads. Forcing all of this into a single scrolling column loses context and makes it hard to see how ideas relate.

The Solution

What if every message was a node, every reply was a connection, and you could see the entire conversation at once?

That's Qonvo. As you chat, a force-directed graph builds in real-time:

Messages → Nodes (color-coded by sender)
Replies → Solid edges (direct connections)
#Tags → Dashed edges (link related ideas across threads)

Click any node to focus it and see its connections. The graph updates live as people chat.

The Tech Stack

Built this as a solo project in about 2 weeks:

Frontend: Vanilla JS + D3.js (no React, no build step)
Backend: Node.js + Express + Socket.io
Persistence: Redis (Upstash) with 3-hour TTL
Hosting: Railway

Why Vanilla JS?

Honestly? Speed. I wanted to ship fast without wrestling with bundlers. The entire frontend is a single 1,800-line HTML file. Is it perfect? No. Does it work? Yes.

D3's force simulation handles the graph physics. Socket.io keeps everyone in sync. That's really it.

The Core Graph Logic

// Every message becomes a node
const node = {
  id: nanoid(12),
  userId: sender.id,
  name: sender.name,
  text: message,
  tags: extractTags(message), // #hashtags
  ts: Date.now()
};

// Replies create edges
if (replyTo) {
  edges.push({ 
    source: replyTo, 
    target: node.id, 
    type: "reply" 
  });
}

// Tags connect related messages
for (const tag of node.tags) {
  const lastWithTag = findLastMessageWithTag(tag);
  if (lastWithTag) {
    edges.push({ 
      source: lastWithTag.id, 
      target: node.id, 
      type: "tag",
      tag 
    });
  }
}

What I Learned

1. Ephemeral is a feature, not a bug

Rooms auto-delete after 3 hours. No accounts, no sign-ups, no data hoarding. Users actually love this — it makes the tool feel safe for quick, throwaway conversations.

2. The graph needs to be optional

On mobile, showing chat + graph doesn't work. So there's a toggle to open the graph as an overlay. Desktop shows both side-by-side.

3. Force simulations are tricky

Getting nodes to not overlap while keeping connected nodes close took a lot of tweaking:

simulation
  .force("charge", d3.forceManyBody().strength(-280))
  .force("collision", d3.forceCollide().radius(45))
  .force("link", d3.forceLink(edges).distance(90));

Try It

Live: qonvo.xyz

There's a demo room with sample data, or create your own room and share the link.

No sign-up. No install. Just click and chat.

What's Next

Paid room extensions — Keep rooms alive for 7/30 days instead of 3 hours
Export to Markdown — Turn conversations into documentation
PWA support — Installable on mobile

Feedback Welcome

This is my first real "ship it and see" project. Would love to hear:

Is the visualization actually useful, or just a gimmick?
What would make you use this over a regular group chat?
Any UX issues I should fix?

Drop a comment or try the demo and let me know what breaks.

Building in public as a solo founder. Follow along: @digi_wares

I Built a Free Article-to-Audio Converter!

Digiwares — Fri, 23 Jan 2026 05:45:01 +0000

I Built a Free Article-to-Audio Converter in a Weekend

My "read later" list was out of control. Hundreds of articles I'd never get to. So I built Sornic — paste any article URL, get audio in seconds.

The Problem

I wanted to catch up on articles while commuting, cooking, or working out. But:

Pocket/Instapaper just pile up unread
Most TTS apps sound robotic
Browser extensions are clunky

I wanted something dead simple: URL in, audio out.

The Stack

Next.js 14 (App Router)
OpenAI TTS API (natural voices)
Vercel (hosting + serverless functions)
Upstash Redis (rate limiting)
Tailwind CSS (styling)

How It Works

1. Article Extraction

When you paste a URL, the server fetches the page and extracts the article content using Mozilla's Readability library (same one Firefox uses for Reader View).

const dom = new JSDOM(html, { url });
const reader = new Readability(dom.window.document);
const article = reader.parse();

For JS-heavy sites that don't work with simple fetch, I fall back to Firecrawl.

2. Content Cleanup

Raw extracted text often includes navigation, ads, "Subscribe now!" prompts. I use Claude Haiku to clean it up:

const response = await anthropic.messages.create({
  model: 'claude-haiku-4-20250514',
  messages: [{
    role: 'user',
    content: `Clean this article for text-to-speech.
              Remove nav, ads, CTAs. Keep only the article body.
              ${rawText}`
  }]
});

3. Text-to-Speech

OpenAI's TTS API has a 4096 character limit, so I chunk long articles:

function splitTextIntoChunks(text: string, maxLength: number): string[] {
  // Break at sentence boundaries when possible
  const sentenceMatch = remaining.slice(0, maxLength).match(/.*[.!?]\s/s);
  // ...
}

Then generate audio for each chunk and concatenate:

for (const chunk of chunks) {
  const mp3Response = await openai.audio.speech.create({
    model: 'tts-1',
    voice: 'nova',
    input: chunk,
    speed: 1.0
  });
  audioBuffers.push(await mp3Response.arrayBuffer());
}

4. Rate Limiting

Free tier = 5 articles/day per IP. Using Upstash Redis:

const ratelimit = new Ratelimit({
  redis: Redis.fromEnv(),
  limiter: Ratelimit.fixedWindow(5, '24h')
});

Challenges

1. Vercel Timeouts

Default is 10 seconds. Long articles can take 30-60 seconds to process. Fixed with:

// vercel.json
{
  "functions": {
    "app/api/**/*.ts": { "maxDuration": 60 }
  }
}

2. ESM/CommonJS Conflicts

jsdom v27 broke on Vercel due to ESM issues. Downgraded to v24:

npm install jsdom@24.1.3

3. Sites Blocking Scraping

Some sites block server-side requests. Firecrawl handles these as a fallback — it uses headless browsers and handles anti-bot measures.

Cost Breakdown

Per article (~2000 words):

OpenAI TTS: ~$0.03
Claude Haiku cleanup: ~$0.001
Vercel/Upstash: Free tier

At 5 free articles/user/day, costs stay manageable with the rate limit.

What's Next

Download as MP3
Browser extension
Playlist/queue feature
Premium tier with more articles

Try It

sornic.com — no signup required.

Drop a URL, pick a voice, hit play. Would love feedback on what features would make it more useful.

Tags: webdev, nextjs, openai, javascript

What Will Actually Work in SaaS in 2026 (And What Won't)

Digiwares — Tue, 20 Jan 2026 09:40:22 +0000

I've been thinking a lot about this lately.

Mostly because of how many projects I've personally built that were GOOD, but still didn't really go anywhere.

70% failed. 30% are in high stormy seas.

They were not broken.
They were not useless.
They were just... optional.

So instead of guessing trends, I started looking at why certain tools quietly disappear and why others stick, even if they're boring.

Here's where I've landed for 2026.

This is opinionated and based on my experience. Happy to be wrong. But this is how I'm planning my own builds now.

What I Think Will Work in 2026

1. Tools That Automate Decisions, Not Just Tasks

Automation today is still very "if this, then that". So many ifs and buts. That's fine, but still shallow.

The tools that matter are the ones that decide:

When something matters
When to escalate
When to block
When to trigger the next step

If your software reduces the number of decisions a human has to make, it becomes hard to remove.

If it only saves a few clicks, it's optional.

That's the math.

2. Software That Lives Inside Real Workflows

Standalone tools are dying quietly.

The stuff that sticks:

Plugs into existing systems
Reacts automatically
Runs in the background
Doesn't require people to remember to open it

If your product depends on someone logging in regularly to "check" something, you're fighting human nature.

3. Boring, But Regulated and High-Stakes Tools

This took me a while to accept.

The exciting ideas are crowded. The boring ones are ignored.

Compliance, finance, ops, contracts, security, governance... these are annoying problems, but they don't go away.

Even though I kept realizing this, I ignored it because I'm not a subject matter expert in these areas.

But people happily pay for software that:

Reduces risk
Avoids mistakes
Prevents "oh shit" moments

Even if the UI is dull.

4. Systems That Get Better The Longer They Run

Most tools reset every session.

The ones that win will:

Remember past decisions
Learn from overrides
Adapt based on outcomes
Improve without constant rebuilding

A perfect example: LLMs like ChatGPT and Claude become better every day.

Software without memory doesn't compound.

5. AI as Infrastructure, Not the Headline

AI will be everywhere by default.

The tools that survive won't market themselves as "AI-powered". They'll just quietly use AI where rules fall short.

Here's the test:

If removing AI breaks your product completely, your tool is a dependency on itself and its existence
If removing AI makes it worse but still functional, that's healthier

The moment your software takes responsibility, it becomes trusted.

What I Think Won't Work (Or Will Be Very Hard)

1. Standalone AI Tools That Only "Generate Outputs"

Text in → text out.

Even if it's good. Even if it's fast. Even if it's cheap.

If it doesn't connect to anything else or drive action, it's a one-time curiosity.

People try it once and move on.

2. Insight-Only Dashboards

I've built these. They look great.

But "here's what's happening" without:

"Here's what we're doing"
"Here's what happens if you don't act"

...doesn't change behavior.

Charts without consequences fade into the background. No one needs them. They get trashed.

3. Tools That Require Constant User Attention

Anything that depends on:

Daily check-ins
Manual reviews
Remembering to come back

...is fragile.

People don't want more software to think about. They want fewer things to worry about.

4. Generic Tools With No Domain Depth

General-purpose tools that are easy to copy and hard to defend.

The deeper the domain, the harder it is to replicate.

"Works for everyone" usually means "perfect for no one."

5. Products That Exist Only Inside Their Own UI

If your product can't:

Trigger workflows
Integrate cleanly
Influence other systems

...it stays isolated.

Isolated tools get replaced. The end.

The Thumb Rule Before Building Your 2026 SaaS

The tools that work in 2026 will:

Automate decisions, not tasks
Operate across systems
Reduce human babysitting
Enforce rules
Earn trust over time

The tools that don't will:

Look impressive
Feel clever
Get tried once
And quietly get trashed

The Bottom Line

Build one tool that fewer people understand at first, but no one wants to remove once it's up and running.

Stop building optional software.

Build essential infrastructure.

P.S. - I'm not saying don't build AI tools. I'm saying don't build tools where AI is the only thing holding it together. Build tools where AI makes an already useful thing even better. That's the difference between a demo and a business.

A small web app where people quietly appear on a world map

Digiwares — Sat, 03 Jan 2026 06:11:42 +0000

I've moved around a lot. Finding information about a new city is easy- Google has everything. But finding people? That's weirdly hard.

So I built MapMates.

The idea

A map. You drop a pin. Others appear. That's it.

No signup walls. No algorithmic feed. No tracking. Just people, quietly appearing on a world map.

Who it's for

Digital nomads wondering who else is in Bali
Expats in Bangkok looking for other remote workers
Freelancers in Lisbon hoping to find a co-working buddy
Retirees in Portugal curious about their community

Tech stack

Next.js 14 (App Router)
Supabase (Postgres + Anonymous Auth)
Leaflet (maps)
Upstash Redis (rate limiting)
Tailwind CSS (styling)
Vercel (hosting)

How anonymous auth works

Most apps force you to sign up before doing anything. I wanted the opposite — drop a pin in 30 seconds, no email required.

Supabase has a neat feature: anonymous authentication. Users get a real auth session without providing credentials. If they clear their browser, they lose access — but for a casual "pin yourself on a map" use case, that's fine.

const { data } = await supabase.auth.signInAnonymously()

One line. User is authenticated. No friction.

Rate limiting

To prevent spam, I added rate limits using Upstash Redis:

1 pin per user (editable)
3 connection requests per day

const ratelimit = new Ratelimit({
  redis: Redis.fromEnv(),
  limiter: Ratelimit.slidingWindow(3, "1 d"),
})

What's next

City pages (mapmates.io/bangkok, mapmates.io/lisbon)
Email notifications when someone wants to connect
Let anyone create their own community map

Try it

👉 mapmates.io

Zoom in. See who's around. Maybe drop a pin.

Would love feedback from the dev community. What would you add?

Building a Browser-Based Voice-to-Text App with the Web Speech API

Digiwares — Fri, 12 Dec 2025 17:49:43 +0000

Building a Browser-Based Voice-to-Text App with the Web Speech API

I recently built a voice-to-text tool that works entirely in the browser — no backend required for the core functionality. Here's what I learned about the Web Speech API and its quirks.

Why Browser-Based?

Privacy is the main sell. Audio never leaves the user's device. No uploads, no storage, no GDPR headaches. For a simple transcription tool, this is a huge advantage.

The Web Speech API Basics

The API is surprisingly simple:

const recognition = new (window.SpeechRecognition || window.webkitSpeechRecognition)();

recognition.continuous = true;
recognition.interimResults = true;
recognition.lang = 'en-US';

recognition.onresult = (event) => {
  const transcript = Array.from(event.results)
    .map(result => result[0].transcript)
    .join('');
  console.log(transcript);
};

recognition.start();

That's it. You now have live speech-to-text.

The Gotchas Nobody Warns You About

1. Browser support is inconsistent

Chrome uses Google's servers (ironically, not fully local). Safari uses on-device processing. Firefox support is limited. Always check:

if (!('SpeechRecognition' in window || 'webkitSpeechRecognition' in window)) {
  // Show fallback UI
}

2. It stops listening randomly

The API has a habit of stopping after silence. You need to restart it:

recognition.onend = () => {
  if (shouldKeepListening) {
    recognition.start();
  }
};

3. Punctuation doesn't exist

The API returns raw words with no periods, commas, or capitalization. You'll need to handle this yourself:

function addAutoPunctuation(text) {
  // Add period after pause patterns
  // Capitalize after periods
  // Handle common patterns like "question mark" → "?"
}

4. Language switching is manual

You need to build your own language selector and set recognition.lang accordingly. The API supports 100+ languages but won't auto-detect.

When to NOT Use Web Speech API

For anything beyond basic dictation, you'll hit walls:

Audio file transcription — API only does live mic input
Speaker identification — Not supported
Timestamps — Not provided
Accuracy requirements — Enterprise use cases need Whisper, AssemblyAI, or Deepgram

I ended up building a hybrid: free tier uses Web Speech API for live dictation, Pro tier uses Whisper for file uploads and higher accuracy.

Native Language SEO Bonus

One unexpected win: I built language-specific pages with native script UI. The Hindi page is actually in Hindi (हिंदी में वॉइस टू टेक्स्ट), not just "Hindi Voice to Text" in English.

Result: Started ranking for native-language searches with way less competition than English keywords.

Try It

I built this into voicetotextonline.com — free to use, no signup for basic transcription.

If you're building something similar, happy to answer questions in the comments.

I built an AI radio that reads the internet to you

Digiwares — Sun, 30 Nov 2025 07:22:46 +0000

Hey folks 👋

I got tired of endlessly scrolling Hacker News during my commute — I love the discussions, but never have enough time (or attention) to read everything.

So I built something small but useful: Tera.fm

👉 https://www.tera.fm

It fetches the latest top stories from places like Hacker News, summarizes them using AI, extracts key viewpoints from the comment section, and reads everything aloud in a natural voice.

Basically:

Hacker News → summarized → spoken like a radio station.

Why I built it
I take a lot of crowded trains and often cook while listening to podcasts. I always wished there was a simple “internet radio” that reads technical news to me without needing to scroll or tap.

No login. No tracking. No feeds to manage.

Just hit play and listen — while commuting, cooking, or at the gym.

What’s inside right now

HN Top
HN New
HN Ask
HN Show

What’s coming next

Reddit (World News, Science, TIL)
Product Hunt
Dev.to trending
GitHub Trending
Custom channels

Stack

Next.js + Vercel
OpenAI Realtime TTS (Alloy voice)
Hacker News API
Clean, minimal UI

Feedback welcome 🙏
This is still early. Would love thoughts on:

the voice quality
summaries
channel ideas
UX improvements

If you listen to technical content or browse HN regularly, I’d love your feedback.

Thanks for reading!

Digiwares