Transcript Speaker Detection isn't perfect

[wesbos]

[wesbos]

still an issue: https://twitter.com/KrisTemmerman/status/1716507884656427469

[wesbos]

Scott is an announcer on some of them. Likely related to the regex

[themisterholliday]

If we have a way to populate the transcript data locally, I could track down these issues.

[wesbos]

Some more details: #1562 (comment)

@themisterholliday I think I can get you a DB dump if you are still interested?

[themisterholliday]

Some more details: #1562 (comment)

@themisterholliday I think I can get you a DB dump if you are still interested?

Yep I'll take a look if you can grab that 👍

[wesbos]

Emailed ya. Some details:

Here is where we actually append the speaker names:

website/src/server/transcripts/utils.ts

Line 112 in 14ecf7d

const sayings: [string | RegExp, string][] = [

And here we filter the flaggings out (less of an issue)

website/src/lib/transcript/Transcript.svelte

Line 23 in 14ecf7d

const scott = new RegExp(/purple cheese before meeting/gi);

[themisterholliday]

Got it 👍
I'll take a look at this and see what i can find

[themisterholliday]

So, I'll break this into three issues:

1. The flags for speaker detection are sticking around in the transcript view
2. Wes or Scott is missing in the entire transcript
3. Scott is mislabeled as Announcer

The flags for speaker detection are sticking around in the transcript view

This can be seen here: https://syntax.fm/show/683/spooky-coding-horror-stories-2023-part-1/transcript
This is because the transcript attributes "My name is Wes. My dog eats food on" to Wes and "the moon." to Scott, which breaks the Regex.

To fix this:

• The Regex could be even more relaxed
• A search for "startsWith" could be added the same as the line for Scott
• or some change could be made to the ingest of transcripts as they are saved to the DB.

I see the first two as still a little "hacky," but getting this right for all occasions seems complicated.

Wes or Scott is missing in the entire transcript

This issue is because speakers are mislabeled (probably while saving the transcript) with "99" as their speaker id.
Then we filter speakers with the "99" id:

website/src/lib/transcript/Transcript.svelte

Line 20 in 14ecf7d

.filter((utterance) => utterance.speakerId !== 99)

If we don't filter, the speakers still have names, so they show up just fine in the recent shows.

But I'm assuming this was causing an issue on some other shows, so if we have those, I can double-check the filter.
On top of removing the filter, we could check for no speaker name, have the entry in the transcript, and label it as "unknown."

Examples:
https://syntax.fm/show/726/is-htmx-a-joke/transcript

• Scott has a speakerId of 99 and is filtered out completely

https://syntax.fm/show/727/how-to-code-opinionated-typescript-stack-tooling-choices-explained/transcript

• Wes has a speakerId of 99 and is filtered out completely

Scott is mislabeled as Announcer

Can you provide the show number we were seeing this? I can't find one, but I'm checking a limited subset.

[wesbos]

sweet thanks. The speaker ID of 99 is important, - I forget why though. Ill check tomorrow.

I think all of these issues are due to the regex either being too relaxed, or not relaxed enough.

I'd have to check, but I don't think I'm saving the speaker's name in the DB, just the speakers number. The problem with our transcript provider is they don't tell you who is 1 or 2, so we have to do that ourselves.

[themisterholliday]

If I'm following correctly the speaker name is correctly found here (and above):

website/src/server/transcripts/utils.ts

Line 49 in 14ecf7d

const speakerName = speakerNames.get(utterance.speakerId);

Which accounts for any speaker id in conjunction with detectSpeakerNames.
Since the speaker id is saved in the DB, in show 727 your id is 99 instead of 1 or 2. (I think the announcer may be 99 in some older shows?)

Ah yea I remember y'all saying the transcript provider doesn't give the speaker which is why this code is required.