3,800+ lectures, morning walks, room conversations, and interviews — with full transcripts, audio playback, and upcoming AI-powered word-level sync. Hear and read Srila Prabhupada simultaneously.
Every lecture transcribed from the original VedaBase 2025 export. Speaker markers, Sanskrit terms in italic, scripture references linked. Meticulously cleaned up over years — still a work in progress.
MP3 audio for every recording hosted at media.prabhupada.io. Playable inline with the transcript. Speed control, position memory, chapter-based audiobook queues.
Every lecture tagged: date, location, type (BG class, SB class, morning walk, room conversation, initiation, arrival address), speaker markers, scripture references.
21-test formatting suite validating markdown structure. Fixing italic pairing, diacritical marks, speaker markers, wiki link formatting. 3,800 files, thousands of issues being resolved systematically.
The next major step: matching every word in the transcript to its exact moment in the audio. This enables highlighted-as-spoken reading, quotable audio clips, and searchable audio.
Run each audio file through a speech-to-text model (Whisper or Gemini) that produces a timestamped transcript. This gives us word-level timecodes — but the text won't match our cleaned transcripts perfectly.
Output: raw AI transcript with timestamps per word/phrase
Our existing transcripts (from VedaBase, manually cleaned) are the source of truth for text. The AI transcript is the source of truth for timing. We align the two using sequence matching — transferring timestamps from the AI output onto our verified text.
This is the critical step: the AI might hear "Krishna" where our transcript has "Krsna" — the alignment handles these mismatches.
Produce standard SRT subtitle files for every lecture. Each subtitle entry maps a passage of text to a time range. These are universal — usable in any media player, embeddable in web players, parseable by apps.
Spot-check alignment accuracy. Flag sections where audio quality is poor (early recordings, background noise, multiple speakers talking over each other). Mark confidence levels per segment. Human review for flagged sections.
Once we have SRT files, the possibilities open up:
Current passage lights up as audio plays, like a karaoke for lectures
Select a passage in the transcript, get a shareable audio clip of just that quote
Search for a phrase, jump to the exact moment in the audio where it's spoken
Compare AI hearing vs existing transcript to catch transcription errors
Transcripts exported from VedaBase 2025 — the authoritative source. Raw text with encoding issues, formatting inconsistencies, and legacy markup.
Years of meticulous cleanup: fixing character encoding, restoring Sanskrit diacriticals, structuring speaker markers, linking scripture references with wiki links, formatting stage directions.
Comprehensive 21-test formatting suite scanning all 3,800 files. Catches unpaired asterisks, broken italic spans, orphaned markers, misplaced speaker names. Thousands of issues identified and being resolved.
AI-powered audio sync: Whisper/Gemini transcription, alignment against verified text, SRT generation, highlight-as-spoken integration in VaniReader and prabhupada.io.
Every lecture is available now — with transcript and audio. Audio sync is the next step.
Browse Lectures at prabhupada.io