Audio/sync issues with Modern Greek

[nolandan]

Hi there,

Excellent job on the app.

I've been writing my own greek learning app, and already had a Greek reader (very similar to what you have in GBT) - and you inspired me to add the audio.

I didn't want to 'steal' your verse timings as I wasn't sure whether you'd be happy about that, so I have calculated my own using your isochron library, and I have discovered that you have a few errors in your own timings that you might want to correct:

(all for the modern Greek pronunciation)

1. Acts 13, v32-33

2. 2 Cor 8 v13-14

3. 1 Jn 1v9-10 - this is actually an audio bug which needs fixing. There's a phrase repeated.

If you're interested, I'm also using whisper to do speech-to-text to create pins for the isochron to improve the quality of verse timings - although your timings are already pretty good.

If I discover any more, I'll let you know!

[nolandan]

4. 2 Thess 2 v 5 - is repeated in the audio
5. Hebrews 1 v 2 - starts late.
6. Mark 6 v 28 - starts early
7. Eph 3 v 18 - starts early
8. 2 Cor 1v7 - starts late (although v6 has a big textual variant, so quite hard tracking this one!)

[nolandan]

9. 2 Tim 2 on the audio seemed to give me 2 Thess 2 - not sure if that was a one off error

[nolandan]

Updated list of timing errors in timings_tk.csv

I've continued refining my automated verse timing pipeline and have now systematically compared against timings_tk.csv across the full NT. Here are all the boundary errors I've found (verified by manual listening):

Verse boundary errors

Each entry is a boundary between two verses where the timing split is misplaced:

• Matthew 10:2/3 — boundary placed too late
• Matthew 15:5/6 — boundary placed too late
• Matthew 15:13/14 — boundary placed at wrong end of a long silence gap
• Matthew 26:56/57 — boundary placed at wrong end of a long silence gap
• Matthew 26:60/61 — boundary placed too late
• Mark 6:27/28 — boundary misplaced
• John 3:35/36 — boundary misplaced
• John 9:8/9 — boundary misplaced
• John 14:5/6 — boundary misplaced
• Acts 13:32/33 — boundary placed after the next pause (reported previously)
• 1 Corinthians 4:9/10 — boundary misplaced
• 2 Corinthians 1:6/7 — textual variant makes this boundary ambiguous
• 2 Corinthians 8:13/14 — boundary placed after the next pause (reported previously)
• 2 Corinthians 12:10/11 — boundary misplaced
• Ephesians 3:17/18 — boundary misplaced
• Philemon 1:11/12 — boundary misplaced
• Hebrews 1:1/2 — boundary misplaced

Total: 17 boundary errors across the NT.

Audio defects in the MP3 files

• 1 John 1:9 — the MP3 repeats a phrase (audio aberration)
• 2 Thessalonians 2:5 — v5 is repeated in the audio. True v6 starts at 0:59

App issue

• 2 Timothy 2 — the GBT app appears to play the wrong chapter audio (plays 2 Thessalonians 2 instead of 2 Timothy 2)

[ethnosdev]

Hi, sorry for my late reply. I've been taking a break for the last couple of weeks.

Thank you very much for finding all of those timing errors! That is super helpful. All of the work we do ourselves is released to the public domain (Matt 10:8), so there would have been no problem at all if you had just "stolen" the current timings, but your providing the improved timings was even better. I'll work on updating those in the next release.

I'm interested in learning more about how you use Whisper in generating the audio timings. I also work on other audio projects that need timings in other languages.

[nolandan]

Hi there,

Thanks for your response.

I was finding with isochron that some of the verse boundaries were unreliable, and I wanted a programmatic way of testing whether they were likely to be correct.

First I attempted silence analysis (to be honest I got Claude Code to do this so I'm not 100% of how it did it!), but I found that this was not 100% reliable either. So I tried to start generating quality/confidence indicators to cross reference with isochron output.

Then I thought what I really wanted was to do a speech->text so I can pinpoint precise words and their timings. This is where Whisper came in. It could recognise about 60% of the words.

I tried to use this to generate pins for isochron. This required de-duping a bit (no point using words that appear more than once in a short sequence) and also only using short common words in short sequences rather than by themselves.

Again I used this to generate a confidence indicator which i combined the silence/isochron output.

Finally by cross referencing my timings with yours, i had a fairly decent list of suspected errors which i then manually verified.

I'll be honest that i pretty much asked AI to do all the coding and cross referencing, but the output was pretty good.

By the way, I have also now trained a speech synthesis model with the same voice meaning I can regenerate words, phrases etc at will. An example output I did was of 1 Clement 1, which of course is not in the NT at all!

The main purpose of that is to allow the users of my app to hear the pronunciation of single words as they answer questions.

I need to check with the audio creator that he's happy for me to use this, but if so, I think this will be a good resource too which could be shared.

[Jim-217]

By the way, I have also now trained a speech synthesis model with the same voice meaning I can regenerate words, phrases etc at will. An example output I did was of 1 Clement 1, which of course is not in the NT at all!

That sounds amazing. Could this be used, for example, for redubbing the existing GNT audio from textus receptus text to SRGNT text?

The main purpose of that is to allow the users of my app to hear the pronunciation of single words as they answer questions.

Would you mind sharing the name of your app? I can’t find it on your github.