How the AI narration works

1. One published article becomes one short script

Articles arrive from the ten publishers' public pages through a crawler that runs roughly every 45 minutes. For each new article a Google Gemini text model writes a Malayalam script of two or three spoken sentences under a fact-constrained prompt: only what the article says, no added numbers, names, quotes, causes or outcomes, and the publisher's name as the sign-off.

• Structured output is validated before it is used: empty, overlong, malformed or non-Malayalam scripts are rejected and never reach the air.
• If every model attempt fails, an extract of the article's own summary is used rather than anything invented.
• Publishers are rotated so one outlet does not dominate consecutive bulletins.

2. A synthetic Malayalam voice, generated once

The script is spoken by Google Gemini audio models. Each bulletin is synthesised exactly once, encoded to a small mono MP3 and stored; every listener hears the same recording, so a popular bulletin never triggers a second generation.

• Audio is served over ordinary HTTP with range requests and caching, so replay and seeking are instant.
• Bulletins run about ten to twenty seconds each. The live stage queues the newest ones first and keeps playing until you pause.
• No microphone, no music beds, no advertising, no video.

3. Transcript and traceability on every bulletin

Every bulletin page shows the spoken script as a transcript with approximate timestamps, the publisher's name, the time it aired, replay audio while the recording is kept, and a direct link to the original article. The stage shows the same text as captions while a bulletin plays.

Questions people ask

Is NewsTeller a news agency or a newsroom?

No. NewsTeller does not report, verify or break news. It reads out short Malayalam summaries of articles that ten Kerala publishers have already published, and names the publisher in every bulletin. The publisher's article is the record.

Who writes the bulletins?

A language model (Google Gemini) writes each script from one published article under a strict prompt: only facts in the article, no added numbers, names, quotes or speculation, short spoken sentences, and the publisher's name at the end. The output is validated before it is voiced; if generation fails, an extract of the article's own summary is used instead.

Is the voice a real person?

No. Every bulletin is spoken by a synthetic Malayalam voice from Google Gemini audio models. It can mispronounce names and places.

Which publishers are cited?

Manorama Online, Mathrubhumi, Asianet News Malayalam, 24 News, MediaOne, Madhyamam, Kerala Kaumudi, Reporter, Kairali News, Janmabhumi. Each bulletin is built from one article by one of them and links to it.

How current is the channel?

New articles are collected from the publishers roughly every 45 minutes and narrated within minutes of arriving. Completed bulletins stay replayable for a rolling window of the most recent recordings; their transcripts and pages stay available afterwards.

Can I read instead of listen?

Yes. Every bulletin page carries the full transcript, the publisher's name, the time it aired and a link to the original article. The live stage also shows captions while a bulletin plays.

What if a bulletin is wrong?

Trust the publisher's article. Bulletins are machine-written summaries and can be wrong in wording, emphasis or pronunciation. The link to the original report is on every bulletin.

Can I subscribe, or use the bulletins elsewhere?

There is an RSS podcast feed at /feed.xml with MP3 audio and transcripts, a sitemap, and a machine-readable summary of the site at /llms.txt. Quoting a bulletin is welcome with attribution to NewsTeller and the cited publisher.