Whisper vs Youtube Transcript API

Abstract

comparing Youtube Transcript API vs OpenAI Whisper in Speech Recognition accuracy

introduction

For the task of video/ audio understanding, and if that video’s available on Youtube, it seems pretty logical to just use the zero-cost Youtube Transcript right?

In this post, we do a quality comparsion between Youtube Transcript and Whisper’s outputs.

To use Whisper, see the previous post here.

For Youtube Transcript, we’ll install the unofficial youtube-transcript-api that everybody uses: pip install youtube-transcript-api

Example Youtube Video

the source used to demo the STT capability is this video taking about DeepSeekV3:

using youtube-transcript-api

very simple and fast, this CLI command will show the transcript for the last 75 seconds of the video:

youtube_transcript_api NJljq429cGk --languages en --format text | tail -n 24

which look like this:

to do this which essentially just means that these models are becoming more commoditized than ever as that progress the technological progress stalls it’s all about what can you build with the existing Frontier models and this is a perfect example of what a company that doesn’t have to have built all the architecture doesn’t need those high-end GPS is able to do simply by taking existing out it also raises questions between open and closed Source models right they were able to do so this is an open source models that was always the fear here with meta’s llama is that it gives the Chinese an advantage it turns out the Chinese didn’t even need llama they just needed stuff that was already put out there by chat GPT and others now wow dear I feel like this is going to be a huge story in 2025 and the geopolitical ramifications of all of this as well uh really appreciate it dear jaosa

using whisper

Two steps…

first we need to download the audio with pytubefix:

pytubefix "https://www.youtube.com/watch?v=NJljq429cGk" -a

then a simple CLI command to transcrible the downloaded audio file (audio.m4a), which will produce a list of transcripts in different formats in the output_dir:

whisper downloaded/audio.m4a --output_dir whisper_transcript/
tail -n10 whisper_transcript/audio.txt

the output looks like:

is a dumbed-down version of the H100s. They’ve been able to do this, which essentially just means that these models are becoming more commoditized than ever. As that progress, the technological progress stalls, it’s all about what can you build with the existing frontier models. And this is a perfect example of what a company that doesn’t have to have built all the architecture, doesn’t need those high-end GPs, is able to do simply by taking existing outputs. It also raises questions between open and closed source models, right? They were able to do so. This is an open source model. That was always the fear here with Meta’s llama, is that it gives the Chinese an advantage. It turns out the Chinese didn’t even need llama. They just needed stuff that was already put out there by ChatGPT and others. Now, wow, Deirdre, I feel like this is going to be a huge story in 2025 and the geopolitical ramifications of all of this as well. Really appreciate it. Deirdre Bosa.

Conclusion

reading both transcripts, whisper is the clear winner; however, quality here comes at a cost of local inference time. So if you are picky about quality and have only a few videos to analyze, use whisper. If you worried about performance and processing a large number of videos, go for youtube-transcript-api.

Just note that whisper can do a lot more tricks as discussed here.