Blog

    The stream started. You were not at the desk.

    Every guide to transcribing a live stream begins with 'paste the URL of the stream'. That instruction quietly assumes the hardest part is already solved: that you knew it started. For the broadcasts actually worth listening to, you usually don't.

    · 7 min read

    Twenty seconds of the sequence: the channel drops, the alert fires, and the transcript is already several minutes deep before anyone opens it.

    The schedule is the least reliable part of the event

    A rally listed for 7pm starts at 7:40. A press conference gets bolted onto the end of something else and appears with no notice. An earnings call runs long and the Q&A - the part with the unscripted answers - lands forty minutes after the part you set a reminder for. None of this is unusual. It is the normal behaviour of live broadcasts, and it is why "start the transcript when it begins" is advice that fails on contact with a real event.

    The failure is not that you miss the whole thing. It is that you miss the opening, and the opening is disproportionately where the tradeable content sits. Prepared remarks come first. The unhedged line usually arrives before anyone has settled into watching.

    Whatever was said before the capture started was never transcribed - not delayed, not queued, simply never heard. Transcription of a live stream is a recording process, not a retrieval one. There is no archive to fetch the first ten minutes from, because for a stream still airing, nobody has written them down yet. That is the same reason ordinary YouTube transcript tools fail on a stream that is still live: the caption file they fetch does not exist until the broadcast ends.

    Why the bell icon does not solve this

    The obvious answer is to subscribe and turn on notifications. Two problems, and only one of them is about reliability.

    The reliability problem is well documented: YouTube's bell reaches a fraction of subscribers, is batched unpredictably, and competes with everything else in a notification tray. Plenty of people have the bell on and still find out from somewhere else.

    The second problem is the one that actually matters here, and it survives even if the bell were perfect. A notification tells you; it does not do anything. You still have to see it, open the stream, paste it somewhere, and start a capture. Best case that is ninety seconds of untranscribed audio. Realistic case is that the ping arrives while you are in a meeting and the answer is twenty minutes.

    A notification optimises for you finding out. What you want is for the work to have started without you.

    Watch the channel, not the video

    The trick is that you cannot watch a video that does not exist yet. A stream announced ten minutes before it starts has no URL to paste in advance, so any workflow keyed to a video URL is structurally too late.

    A channel, on the other hand, exists right now and is stable. Channel drop alerts key off that: you arm an alert against a channel, LiveScript polls its feed every 30 seconds, and when something new appears the alert fires. You are not predicting the video. You are standing at the door it has to come through.

    Two design details are worth knowing because they change how you use it:

    • It baselines when you arm it. LiveScript records the channel's newest video at that moment and only fires on entries above it, so arming an alert never triggers a backlog of emails about last week's uploads.
    • It fires once, then disarms. An alert that caught a drop switches itself off rather than continuing forever. That is a deliberate safeguard - an unattended watcher cannot quietly start transcripts for months - and it means re-arming is a thing you do on purpose after each catch.

    What “auto-transcribe” actually starts

    Leaving Auto-transcribe when it drops ticked is what turns the alert from a notification into work already done. What it starts depends on what dropped, and the two paths are genuinely different pipelines:

    • A live stream goes to the real-time capture pipeline. Audio is transcribed as it airs, so the transcript grows while the broadcast is still happening - and it started at the drop, not at whatever time you opened the email.
    • A regular upload has no live audio to capture, so it runs a backfill instead: YouTube's own captions first where they exist, then audio transcription for the rest. Same outcome from your side, a transcript waiting when you arrive.

    Both run on LiveScript's servers, which is the part that makes this useful overnight. Nothing depends on a browser tab staying open or a laptop staying awake. The alert is armed server-side, the poll runs server-side, and the transcription runs server-side. Your machine can be off.

    If you only want the heads-up and would rather decide for yourself, untick the box. The email still arrives; it just links to a stream nobody has started transcribing.

    Setting one up

    Three of the five steps happen before anything airs, which is the whole point.

    1. Paste the channel

    Open LiveScript, go to Drop alerts, and paste the channel URL or @handle of any public YouTube channel. One alert per channel per account - arming a second one for the same channel is rejected rather than silently duplicated.

    2. Leave auto-transcribe on

    It is on by default. This is the setting that decides whether you get a link or a transcript.

    3. Walk away

    There is no step here. That is the feature. The most common mistake is treating a drop alert like a calendar reminder and hanging around for it.

    4. Open the email

    It arrives within about a minute of the video appearing, names the channel and the title, and links straight through to the transcript on the watch page. If auto-transcribe was on, the transcript is already several minutes deep by the time you read the subject line.

    5. Tag, then re-arm

    Add the phrases you care about to the running transcript so you get pinged on the words rather than reading everything - the full walkthrough is in monitoring live streams for spoken keywords. Then re-arm the channel alert so it is ready for the next drop.

    Worth being straight about a limit here: keywords attach to a transcript that exists, so for the stretch between the drop and you opening the email, you are reading back rather than being pinged live. That stretch is transcribed and searchable - which is the entire difference from not having armed the alert - but the alerting on it is retrospective. Tag first, then scroll up.

    Where this is worth the setup

    The pattern pays off wherever the start time is unreliable and the opening matters:

    • Mention markets on unscheduled broadcasts. A contract settling on one spoken word is worthless to you if the word was said before you started listening.
    • Political channels with no fixed schedule. Rallies, unannounced addresses, and pressers appended to other events.
    • Overlapping events. Two things at once means one of them gets watched. Alerts and auto-transcription mean neither gets missed, and you read both afterwards.
    • Anything in an inconvenient timezone. The transcript of a 3am broadcast is worth as much at 8am as it was live - as long as somebody was recording it. Nobody was, unless you arranged for it.

    If you would rather this lived in code

    Drop alerts route through email and a person. If your own system should react instead, the live WebSocket API streams transcript lines to your program as they are spoken, so your matching logic runs against text rather than your reaction time. Endpoints and event shapes are in the API reference, and the MCP server covers the after-the-fact half - handing a finished transcript to Claude or another MCP client once the broadcast is over.

    The short version

    Live transcription is a recording process, so the only minutes you get are the ones something was listening for. Every workflow that starts with "paste the stream URL" hands that responsibility to a human who has to be awake, at the desk, and correct about the start time.

    Watching the channel instead moves the trigger to the one thing that reliably exists before the event does. Arm an alert on a channel you care about and the next time it goes live unannounced, the interesting question is what was said - not whether anyone caught it.

    Frequently asked questions

    Why not just use YouTube's bell notification?

    The bell tells you something happened; it does not start anything. Even when it arrives on time, you still have to open the stream and start a transcript yourself, which means the opening minutes go untranscribed. The bell is also famously unreliable - it reaches a fraction of subscribers and is easy to lose in a crowded feed.

    How quickly does the alert fire after a stream starts?

    LiveScript polls each watched channel's feed every 30 seconds, so the alert typically reaches you within a minute of the video appearing. It is not instant and does not claim to be - the comparison that matters is against noticing an unannounced stream yourself, which can take an hour or never happen at all.

    Does auto-transcribe work on regular uploads too, or only live streams?

    Both, through two different paths. A live stream goes to the real-time capture pipeline and transcribes as it airs. A regular upload has no live audio to capture, so it runs the backfill instead - YouTube's own captions first, then audio transcription. Either way you open the email to a transcript that already exists.

    Do I need to leave my computer on?

    No. The watcher runs on LiveScript's servers, not in your browser. Once the alert is armed you can close the tab, shut the laptop, and still get the email - and the transcript will have been running the whole time.

    What happens to videos the channel posted before I set the alert?

    Nothing. When you arm an alert, LiveScript records the channel's current newest video as a baseline and only fires on entries that appear above it. You will not get a burst of emails about last week's uploads.

    Does the alert keep firing for every future video?

    No - each alert fires once and then switches itself off, so an unattended watcher cannot fill your inbox or start transcripts you did not ask for. Re-arming it is one click, and re-arming resets the baseline to whatever is newest at that moment.

    Can I set keyword alerts in advance, so they are live from the first line?

    Not from the self-serve alerts page today - keywords are added to a transcript once it exists. In practice the transcript is searchable from its first line, so the words spoken before you tagged anything are still there to search; you are reading back rather than being pinged live for that opening stretch.

    How many channels can I watch, and what does it cost?

    Add as many channels as you need, one alert per channel. Drop alerts are a Pro feature - $19.99/month, or a $4.99 Day Pass, which also covers the monthly minute pool, unlimited concurrent streams, keyword alerts, multilingual mode, and VTT/SRT export.

    Transcribe a stream that's live right now

    Paste the URL of any live YouTube or Twitch stream and watch the transcript build in real time. Free plan, no credit card.