Better solution out of the box: whisper

When your GPU allows it, you may consider using openai's whisper to convert audio to text
As I did with medium model

medium 	769 M 	medium.en 	medium 	~5 GB 	~2x    

A few moments to keep in mind

• too big of the file (>~1GB) can freeze my machine (likely memory related)
• video stream is not needed
• using cuda depending code requires a proper gpu (nvidia)

pacman -Q | grep cuda
cuda 11.8.0-1
cuda-tools 11.8.0-1   

getting the audio

ffmpeg -i v13_1.webm -vn -acodec copy v13_1.ogg    

• -i: source file
• -vn: no video
• -acodec copy: copy audio (instead of processing/converting)
• v13_1.ogg: output file

using whisper it shall return three new files (.vtt .srt .txt) in directory that it was called from

whisper --model medium --language de v13_1.ogg

du -h v13_1.ogg*
416M    v13_1.ogg
644K    v13_1.ogg.srt
440K    v13_1.ogg.txt
612K    v13_1.ogg.vtt

And time to compare quality of the result

./quality_check.py overlord_audiobooks/vol_13/v13_1.ogg.txt 2>/dev/null
0.8536507196550318:68644:overlord_audiobooks/vol_13/v13_1.ogg.txt

• image recognition: 0.813456818978614
• audio recognition: 0.8536507196550318

Name problem

Reading the whisper result
Whisper gets confused about how to spell fantasy names
Some of variations of Neia (Baraha):

• Neyyaa
• Nehiaa
• Neyar
• Nehia
• Nehiawas
• Nerea
• Neha
• Neia (oh wow, correct one also used)
• Nehiyah
• Ne'ya (-_-)
• Neyawas
• Nelia