AVSenseAI
Architecture · end to end

One listening. Every output.

AVSenseAI turns any recording, in any Indian language, into a record you can search, click, quote, ask, clip and publish. The video stays on the user's device; Gemini on Google Cloud does the understanding; every line plays the moment it was said.

SearchableClickableQuotableAskableClippablePublishable
0minutes in one recording, tested end to end
~0 sline timing error, down from 8.7 s
0%of lines matched to exact word times
₹0/hrmeasured AI cost on Basic
Market

A growing market, priced for everyone

Published estimates from research firms and industry reports, worldwide and in India. Different firms define these markets differently, so they are shown as ranges with their sources, not added together.

🌍 Worldwide

AI transcriptionMarket.us$4.5B → $19.2B
2024 → 2034, 15.6% a year
Speech and voice recognitionMarketsandMarkets, 2025 → 2030, 19.1% a year$9.7B → $23.1B
AI speech-to-text tools, growthTechnavio, 2024 → 2029, 28.8% a year+$8.3B
Creator economyGoldman Sachs: about 50 million creators, growing 10–20% a year$250B → $480B

🇮🇳 India

Speech and voice recognitionMarketsandMarkets, 2025 → 2030, 21.1% a year$366M → $954M
Vernacular news and contentKen Research, 2025 → 2031$1.35B → $3.1B
Indian-language online news readersKantar with Google, 2023379M
Content creatorsISB–Hashfame, 2025 (0.96M in 2020); 66% outside the metros4.12M
Registered newspapers and periodicalsMinistry of I&B, 2024–251,54,523

What one hour of recording costs today

ToolWhat you getPrice for 1 hour of audio
Human transcription (Rev)transcript$1.99 a minute ≈ $119
AI transcription (Rev Essentials)transcript, monthly plan$29.99 a month for 5,000 AI minutes
Otter Protranscript, English-first, monthly plan$16.99 a month ($8.33 billed yearly) for 1,200 minutes
Google Cloud Speech-to-Textwords only$0.016 a minute ≈ $0.96 (batch $0.003 ≈ $0.18, results within 24 h)
Sarvam speech-to-textwords in Indian languages₹0.49 a minute ≈ ₹29 (₹0.59 ≈ ₹35 with speakers)
AVSenseAI Article onlya finished article, any Indian languageAI cost ≈ $0.02 (measured: 72 min for $0.024)
AVSenseAI Low costwho said what and how + article + AskAI cost ≈ $0.05 (72 min for $0.056)
AVSenseAI Pro Maxexact times, speakers, highlights, clips, article, AskAI cost ≈ $0.27 (18.5 min for $0.082)

Our figures are measured AI cost per recording; the user price adds the margin set in admin. Others are their published list prices, which change over time.

Sources: Market.us AI transcription · MarketsandMarkets speech & voice recognition · MarketsandMarkets India · Technavio AI speech-to-text tools · Goldman Sachs creator economy · Ken Research vernacular news · Kantar–Google Indian digital news consumer 2023 · ISB–Hashfame creator economy · Registered publications 2024–25 · Rev pricing · Otter pricing · Google Speech-to-Text pricing · Sarvam pricing (all checked 21 Nov 2026)

Built on Google

The Google stack

Gemini 3.8 FlashPro Max listening and writing
Gemini 3.5 Flash-LitePro listening and matching; fact check and "tell me more"
Gemini 2.5 Flash-LiteBasic: long recordings at the lowest cost
Vertex AIpro tiers via service account: no busy queue
Gemini API (AI Studio)Basic tier key; prompts built and tested in AI Studio
Cloud RunAPI, worker and word timing in one container; scales to zero
Cloud Build + Artifact Registrybuild from the server, keep the latest 2 images
Cloud Storagethe record database, copied every 5 seconds
Cloud Billing budgetsmonthly alert at 50, 90 and 100%
IAM service accountsleast access: Vertex AI user, bucket writer, builder
Grounding with Google Searchfact check and "tell me more": what published sources say, with real source links
Google Fact Check Tools APIexisting fact-checks of a post or a claim, with the fact-checker's own rating
API keys (restricted)the fact-check key works only for that one API
Search Consoledomain verification for the demo address
Tiny local speech modelopen model for exact word times, runs inside our container, no per-minute bill
Mediabunny + WebCodecssound, clips and stills made in the browser
WordPressthe AVSense News blog on our own server
Litestreamkeeps the database safe in Cloud Storage
Why AVSenseAI

Words alone are not the answer

A reporter with a 1-hour press meet, a teacher with a recorded class, or anyone sent a viral clip in a language they do not speak needs the meaning: who said what, how, and what it means, in their own language. Plain speech-to-text gives only words, at a price per minute, and leaves the rest to the user.

The usual wayWhat you getPrice for 1 hour
Google Cloud Speech-to-Text (standard)words; Tamil was not on its speaker-separation list when we checked$0.016 a minute ≈ $0.96 (batch ≈ $0.18, up to 24 h)
Sarvam speech-to-textwords in Indian languages; speakers cost extra₹0.49 a minute ≈ ₹29 (≈ ₹35 with speakers)
Human transcription (Rev)words, in days, not minutes≈ $119
Then: an AI model or a personto read it all and write the story, notes or answermore cost, more time
AVSenseAIone listening: words, speakers, tone, sounds, then the article, answers and clips, in any languageArticle only ≈ ₹2 · Low cost ≈ ₹5 (measured AI cost)

🗣 Indian speech is hard

Mixed Tamil, Hindi and English in one sentence, phone lines, crowds and cross-talk. A model that understands the meaning handles it far better than one that only matches sounds to words.

🖥 Servers cost even when idle

Hosting your own speech model means a GPU server paid for all month. Our timing model is tiny and runs on CPU inside Cloud Run, which scales to zero: no recording, no bill.

📱 Viral clips, any language

Paste a reel or a WhatsApp video in Telugu or Hindi and read what happened, who said what, and how, in Tamil or English. Only what is heard, never guessed.

Sources: Google Cloud Speech-to-Text pricing · Sarvam API pricing · Rev pricing. Speaker-separation languages from Google's Chirp 3 documentation, checked 11 Nov 2026. AVSenseAI figures are our measured AI cost (72-min Tamil call).

Watch it work

One file in. Everything out.

A real run, replayed: an 18-minute Tamil press meet video goes in, and you can follow it from the device through the backend to every output. The timings shown are from our real test.

① Your device

AVSense desk · New
⬆
Drop a file or paste a link
🎬 dgp-press-meet.mp4 1.2 GB
Taking only the sound…
video 1.2 GBsound 4 MB ✓
🔒 The video stays here. Clips and stills are cut here too.

② Google Cloud backend

⚙
Job receivedCloud Run API queues it for the worker
waiting
🎧
Gemini listenswords, speakers, names, tone, applause
waiting
⏱
Word timing (tiny local model)every word's exact second, at the same time
waiting
🧩
Gemini matches + findsright words on exact times · questions, topics, highlights
waiting
🗄
Saved safelyrecord copied to Cloud Storage
waiting

③ What you get

📝
Clickable transcript03:12DGP: “Law and order is fully under control.”
👥
Who spoke whenthe DGP and the reporters · lanes you can play one by one
✦
Highlights11:40“Zero tolerance” on custodial deaths
❓
Questions + answersevery reporter question with its answer, playable
💬
Ask it anything“What did he say about drugs?” → answer with stills
✍
News article, any languageTamil or English, headline, quotes, key facts
✂
Clipsone line, one answer, or all of one speaker, joined
📰
Publishedlive on AVSense News · Word · PDF · subtitles
00:00Getting ready…
How it flows

From a file or a link to a finished record

Three places do the work: the user's device (privacy and speed), Google Cloud Run (the brain), and our own server (the news blog and demo files). Moving dashes show where data travels.

ON YOUR DEVICE GOOGLE CLOUD RUN · AVSENSE OUTPUTS · OUR SERVER 🎬 File or social linkvideo, audio · any language 🔒 Sound taken on device1.2 GB video → ~4 MB soundvideo never uploaded ▶ Player + ✂ clipsplay any line, cut clips,join one speaker's parts 🖼 Frame stillsfrom your own video ⚙ API + workerjobs, guest sign-in, live status 🎧 Gemini listenswords, speakers, names, tone,laughter, applause, playback ⏱ Tiny local speech modelonly the clock: every word'sexact time, in our container 🧩 Gemini matches + findsright words on exact times (98% of lines)questions · topics · highlights · moments ✍ Write · 💬 Askstreamed, any language out 🗄 Record databasecopied to Cloud Storage every 5 s 🖥 The desktranscript, lanes, tabs,highlights, answers 📰 AVSense Newsone click to the blog ⬇ DownloadsWord, PDF, subtitles, text
stays on the deviceGemini (Google AI)Google Cloudtiming and outputs
Step by step

What happens when you press Go

1

Only the sound travels

The browser takes the sound track out of the video on the device (mono, 16 kHz). A 1.2 GB video becomes about 4 MB, so it reaches us in seconds and the pictures never leave the computer.

2

Two ears at once

Gemini hears the meaning: words, who speaks, names when said, tone, laughter, applause, a clip played in the room. At the same time an open Indian speech model writes down every word with its exact time.

3

Right words on the right second

Gemini places its words on the exact timed lines (text only, in parallel pieces), then finds the questions, topics, highlights, moments and a heard note on tone and mood.

4

Everything clicks

Every line, question, topic and highlight plays its moment in the player. Click a speaker to hear only them; cut a clip, or all of one person's parts joined, right on the device.

5

Ask and write, live

Ask anything and get answers with video stills of the moments; write a news article, notes or a report in any language, streamed as it is written, with optional notes to steer it.

6

Publish in one click

Send the article to the AVSense News blog, or download Word, PDF, subtitles or text. The record is kept safe in Cloud Storage and comes back after any restart.

Design choices

Why we chose this path

Every part of the design answers one question: how do we give a reporter the exact moment, in their language, fast, private and at the lowest cost?

🎧 Gemini is the core

One model hears the meaning, who speaks, names, tone, laughter, applause and a clip played in the room, in any Indian language and mixed speech, and writes in any language.

Instead of: a cloud speech-to-text service plus a separate model. That means two bills, and plain text with no tone or meaning.

⏱ A tiny local model only for the clock

Gemini knows what was said; a tiny open speech model running in our own container knows exactly when, word by word. Together: the right words on the right second.

Instead of: trusting AI line times (up to 8.7 s late) or paying per minute for cloud timing.

👂 Listen once, think many

The costly step, listening to the audio, happens once. Write, Ask, extra articles and translations then work on the text: about a cent per article, a fraction of a cent per answer.

Instead of: sending the audio again for every question.

🔒 Only the sound leaves the device

The browser takes the sound out of the video, so a 1.2 GB video uploads as about 4 MB. It's private, saves mobile data, and starts in seconds. Clips and stills are cut on the device from the user's own copy.

Instead of: uploading and storing everyone's videos on a server.

🧩 Match as text, in parallel pieces

After listening, Gemini places its words on the timed lines as a text-only step, in pieces at the same time. It's fast, cheap, and the words don't drift.

Instead of: asking the model to fix times while hearing the audio (only 15–57% of lines held).

💸 Pay only when used

Cloud Run scales to zero; the record database is a single file copied to Cloud Storage every 5 seconds; Vertex AI for the pro tiers avoids the busy queue. The blog runs free on our own server.

Instead of: an always-on database server and fixed monthly bills.

Recording types

Tuned for 19 kinds of recording

The user picks what the recording is, and the listener is told what to expect: phone sound, cross-talk, crowd noise, a guest introduced by name, a teacher writing on the board.

Phone interviewInterview (in person)Host interview + audience questionsGuest talk at a tech event or hackathonViral or social media clipPress meetPress meet with cross-questioningTV debate with named guestsTV news clipPublic or political meetingSpeechClass or lecturePodcastDiscussion or debateZoom or online meetingHearing or inquiryOn the spot (field)Religious talkPhone call
Output types

38 ready outputs for 6 kinds of user

Each output has its own editor prompt: headline, lead, sub-headings, quotes, key-facts table. All of them can be written in any language, at any length and tone, with the user's own notes.

Journalist 17

What happened in this clipNews articleFeatureInterview Q&ANews briefPress releaseEditorialDebate reportPress meet reportNews analysisKey pointsQuote sheetTimelineLive updatesTV / radio scriptSocial postsFact-check brief

Teacher 8

Lecture notesExam with answer keyRevision sheetWorked solutionsQuestion and answer setPractice sheetFull class packFormula sheet

Tech writer 4

Tech blog postKey takeawaysExplainerSocial thread

Business 3

Meeting minutesAction itemsExecutive brief

Creator 4

What happened in this clipBlog postSocial packShort-video script

Researcher 2

Research reportQuote bank

Transcript, three ways

As spokenAs spoken, with translationTranslated into your language

Classes: 17 subjects

MathsPhysicsChemistryBiologyComputer scienceEngineeringStatisticsEconomicsAccounts and commerceMedicine and healthLawHistoryGeographyPolitical scienceLiteratureLanguage and grammarOther subject

Teacher (school) or professor (college) voice, with maths and chemistry symbols.

Download as

WordPDFSubtitles (SRT)TextMarkdownClips (MP4)Blog post
Hard problems, measured fixes

What broke, and how we solved it

Every fix below came from a test on real recordings: Tamil press meets, a 72-minute phone interview, a maths class, a TV debate, a Telugu news clip.

Long recordings fell apart

One call on a low-cost model broke after about 12 minutes; it looped or stopped early.

Fix: cut at real pauses into about 10-minute pieces, watch the stream for loops and keep what was good.

72 min complete · $0.07

Times drifted seconds late

On a long monologue Gemini's own line times ran up to 8.7 s late, so clicks played the wrong words.

Fix: an open Indian speech model times every word; Gemini only supplies the words and meaning.

8.7 s → about 1 s

Editing while listening drifted

Asked to correct timed lines while hearing the audio, models re-wrote and drifted: only 15 to 57% of lines kept their words.

Fix: listen first, then match words to timed lines as a text-only step.

98% of lines matched

"Model busy" at peak hours

The newest model queued for about 4 minutes or returned busy errors on the plain API key.

Fix: run the pro tiers on Vertex AI with a service account.

first words in 4 s · 18 min in 38 s

Low-cost model ignored formats

It copied template lines and wrote topics without times, so cards could not play.

Fix: structured JSON output for speakers and extras.

0 → 42 moments, 24 topics · $0.003

Speakers changed between pieces

Each 10-minute piece numbered voices on its own.

Fix: every piece is told the speakers found so far; a "more than one person" switch tells the model what to expect.

84/84 and 77/77 lines labelled

Huge video uploads

Reporters upload 1 to 2 GB videos on slow mobile data.

Fix: take the sound on the device; clips and stills are cut on the device too.

1.2 GB → about 4 MB

Waiting felt broken

Users stared at a still screen for minutes.

Fix: stream the listening and the writing; an animated agent explains each step and its benefit.

live text in seconds

Surprise cloud bills

AI and cloud costs can run away in a demo.

Fix: real cost from every response, a monthly budget alert, one instance that scales to zero, and a capped guest credit.

₹900/month alert · $5 guest cap
Tiers

Three tiers, one desk

Measured AI cost per hour of recording on our tests; the user price adds a margin set in admin (now 20%). Writing an extra article costs about one cent; an Ask answer, a fraction of a cent.

Low cost

Gemini 2.5 Flash-Lite (set in admin) · text only
≈ ₹3–5 / hour of audio

Measured: 11-minute Tamil interview, transcript + article $0.009; article only $0.007.

  • Who said what, and how: speakers, tone, sounds
  • No timing model, so any language
  • Article only (every tier): 72 min for $0.024
  • Or transcript first, then the article
  • Ask it anything, Write more, Publish to blog
  • Done in about half a minute for 11 minutes

Basic

Gemini 2.5 Flash-Lite · for long recordings
≈ ₹4 / hour of audio

Measured: 41-minute Tamil call for $0.031 (about $0.045 an hour).

  • Exact word timing, every line clickable
  • Hours-long recordings (72 min tested)
  • News article, notes or report in any language
  • Best 5 topics, highlights, questions, moments
  • How it sounded: tone and mood note
  • Ask, Write, clips, Publish to blog
  • Speakers on request ("more than one person")

Pro

Gemini 3.5 Flash-Lite · on Vertex AI
≈ ₹15–18 / hour of audio

Measured: 4-minute news clip for $0.014; 18-minute class for about $0.05.

  • Everything in Basic
  • Gemini-corrected words on exact times
  • Speakers, names when said, tone on each line
  • A clip played in the room marked as playback
  • All questions, topics, highlights, moments
  • Translated transcript into your language
  • Classes: problems with steps and maths symbols
BEST

Pro Max

Gemini 3.8 Flash · on Vertex AI
≈ ₹24–28 / hour of audio

Measured: 18.5-minute DGP press meet for $0.082; 13.8-minute TV debate for $0.074.

  • Everything in Pro
  • Best names from introductions ("over to you, Sundar")
  • Mixed languages in one speech (Tamil, Hindi, English)
  • Reporter cross-questioning, heated moments
  • Write while listening: tone straight into the article
  • Best for press meets, debates, town halls
Full feature list

Everything a recording becomes

Any language in and outIndian languages and mixed speech in; read or write in your language
Paste a social linkor drop a file; only the sound is taken
Clickable transcriptevery line plays its moment; read along as it plays
Who spoke whenspeaker lanes, names, roles; rename with one click
Play one speakerhear only one person, part after part
✦ Highlightsthe moments not to miss, with why they matter
Questions + answersevery question with its answer, playable
Topicsthe recording split by topic
Momentsapplause, laughter, raised voices, key lines
How it soundeda reporter's note on tone, mood and the room
Ask it anythingstreamed answers with video stills of each moment
Write anythingnews, analysis, debate report, quotes, timeline, TV script, social posts, notes, worked solutions
Your notessteer a piece: focus, length, tone
Clips on devicecut a line, a question, or all of one speaker, with a mark
Classes17 subjects, teacher or professor voice, maths and chemistry symbols
Publish to blogone click to AVSense News
Article onlystraight from the audio on any tier; length follows the recording (72 min → $0.024)
Join up to 5 filesparts of one recording, joined in your order, heard as one
Fact check + tell me moreexisting fact-checks and what published sources say, for any recording or pasted post; no opinion added
Low cost tierwho said what and how, without times, any language, Ask with questions from the recording
DownloadsWord, PDF, subtitles (SRT), text, Markdown
Real cost showntime taken and cost for every recording
Live agent statusan animated guide explains each step
Admin controlmodels, engines, prices, prompts, save, publish, restore