If someone is interested, this is my supershort zsh/bash scripts that I keep in .zshrc for doing the same thing using plain whisper.cpp, ffmpeg and yt-dlp (`brew install whisper-cpp yt-dlp` for Mac); I output it in vtt format (subtitles) though, but it's easy enough to change it to txt.
yt_to_srt() {
local url="$1"
local output_base="$2"
local language="${3:-en}"
yt-dlp -x --audio-format wav --postprocessor-args "-ar 16000" -o "$output_base.wav" "$url"
whisper-cli --language "$language" --model "$WHISPER_MODEL" --split-on-word --max-len 65 --output-vtt --output-file "$output_base" --file "$output_base.wav"
rm "$output_base.wav"
}
file_to_srt() {
local filepath="$1"
local language="${2:-en}"
local filename=$(basename "$filepath")
local filename_no_ext="${filename%.*}"
local output_base="$filename_no_ext"
local temp_wav="$output_base.wav"
ffmpeg -i "$filepath" -vn -acodec pcm_s16le -ar 16000 -ac 1 "$temp_wav"
whisper-cli --language "$language" --model "$WHISPER_MODEL" --split-on-word --max-len 65 --output-vtt --output-file "$output_base" --file "$temp_wav"
rm "$temp_wav"
}
plus additional bootstrap script for large-v3-turbo model from my chez-moi dotfiles:
#!/bin/bash
# Download whisper.cpp models from Hugging Face (runs once per machine).
set -euo pipefail
MODELS_DIR="$HOME/whisper-models"
BASE_URL="https://huggingface.co/ggerganov/whisper.cpp/resolve/main"
MODELS=("ggml-large-v3-turbo.bin" "ggml-tiny.bin")
mkdir -p "$MODELS_DIR"
for model in "${MODELS[@]}"; do
if [ ! -f "$MODELS_DIR/$model" ]; then
echo "Downloading $model..."
curl -L --progress-bar -o "$MODELS_DIR/$model" "$BASE_URL/$model"
else
echo "$model already exists, skipping."
fi
done
echo "Whisper models ready at $MODELS_DIR"
So just tested this on a bunch of videos. It's crazy accurate (not 100% obviously), fast and resource efficient.
Not only was I testing the stt part, but also, timestamps and speaker identifications. Tested on 3 different videos, local and online both.
Timestamps were precise down to sub 500-ms level even on longer 20+ minute videos. Speaker identifications worked equally well. My old M1 Air didn't hang a single bit while the transcription was going on.
Am I a bit thick, but first we created this amazing way to transfer any text very cheaply and fast over network, then we (well, I think it was Meta and Google) decided that no, everything must be a video, then we added subtitles and AI-transcriptions to those videos and now we just dowload transcriptions of those videos presumably to feel LLM to make summaries of them in order to… Read. Them.
Good point! I haven't found a faster way to consume info than reading. But depends on the type of learner you are (visual, auditory, hands-on/interactive, etc)
So, this project consists of a ~175 line README and a ~500 line Python program that glues yt-dlp and Kroko together. Neat.
I guess if it encourages you to install and figure out how to use ffmpeg, yt-dlp, kroko, numpy, and onnx that's a good thing. Sometimes just knowing a thing is possible is a huge benefit.
I see the value as a centralized anti-content-blocker.
This repo is now a good way to centralize hacks around the sure-to-come blockers those platforms will add to prevent download.
Just like uBlockOrigin was a way to centralize all the "just run this greasemonkey script" comments, I can see this getting a huge following for people who really value transcriptions.
thank you. You nailed the actual value, that's right. The real win is just knowing you can do this on a laptop CPU, offline, no GPU or cloud bill. There are tiny done-for-you details, like rescaling token timestamps back to real time after the atempo speedup so --timestamps doesn't lie to you, but they are minor.
Do you have a link to results confirming this? Kroko does not seem to be on the Open ASR Leaderboard. Parakeet has an average WER of 6.32 across several common datasets.
Kroko's website says benchmarks aren't formalized yet. FWIW, this url says 5% WER for English [0]. though it doesn't specify the dataset, so not directly comparable to Parakeet's 6.32 on the Open ASR Leaderboard
Very cool, I'm also working on a captioning/subtitling project for the lecture recordings for the university I work at.
My biggest challenge is finding a proper language model that is fast enough and accurate enough since I have to caption about 600 hours of video per week and I preferably want to run all of this on a tiny server (2 cores 4 GB memory). This tool could easily do that with the kroko model but I'll have to test if the accuracy is good enough.
Also in my own scripts I'm using ffmpeg to download just the audio of the videos that I want to caption, which saves a lot of bandwith and speeds up the whole process. As far as I can see this tool doesn't do that, that would be a nice functionality to add, plus an option to turn the output into a working .srt file.
Tis is very simple and very cool! Just installed it on my Hetzner box where I run a remote controlled local agent so now I can basically chat/email a video link to get a summary and/or ask questions. The only issue was YouTube's PO Token requirement (web/mweb clients refuse to serve formats from datacenter IPs without a valid Proof-of-Origin token.) So I had to find a client that still work without PO Token first. Thanks for sharing!
Had Claude test it out on 3 videos. Worked at 5-8x realtime. The beauty of it is that it works on all videos, not just the one with transcripts. Combine it with YouTube search and LLM takeaways from transcripts, and you have super-efficient content consumption. There are SaaS products that charge 1 cent per video for those with transcripts. There is a viable product in here somewhere, methinks.
Wouldn't it still be more efficient to do GPU transcriptions anyways? is this something we could actually put the effectively useless NPUs to use in modern laptops?
small, ONNX-optimized models designed specifically for low-latency CPU streaming, so it avoids overhead of large transformer arch and GPU memory transfers
Youtube has transcripts on most videos, not all. The others don't expose them. If you mean the "transcript APIs" for TikTok/IG/X, they are all transcribing audio like yapsnap does. If you have a way to pull native ones, let me know, genuinely curious.
Just download the model for your preferred language, all hosted on the Kroko-ASR collection here: https://huggingface.co/Banafo/Kroko-ASR/tree/main
Right now you have Dutch, French, Portuguese, Spanish, German, Italian, Swedish, Swiss German, Hebrew, and Turkish. Grab the one that matches your audio, point yapsnap at it with --model (or set KROKO_MODEL), and you're set!
Added Diarization / Speaker Separation that is fast and CPU only. Thank you all for the great feedback and support. PRs welcome!
yapsnap "https://www.youtube.com/watch?v=NzKJ-xO-VhE" --diarize
SPEAKER_00 [00:00]: Welcome to the show.
SPEAKER_01 [00:03]: Glad to be here, thanks for having me.
SPEAKER_00 [00:08]: Let's get started.
If someone is interested, this is my supershort zsh/bash scripts that I keep in .zshrc for doing the same thing using plain whisper.cpp, ffmpeg and yt-dlp (`brew install whisper-cpp yt-dlp` for Mac); I output it in vtt format (subtitles) though, but it's easy enough to change it to txt.
plus additional bootstrap script for large-v3-turbo model from my chez-moi dotfiles:yt-dlp can download auto-subtitles and regular subtitles, why not do that and fall back to whisper?
To be frank I didn't know there's such an option :-)
In my experience Whisper is several orders of magnitude slower though.
So just tested this on a bunch of videos. It's crazy accurate (not 100% obviously), fast and resource efficient.
Not only was I testing the stt part, but also, timestamps and speaker identifications. Tested on 3 different videos, local and online both.
Timestamps were precise down to sub 500-ms level even on longer 20+ minute videos. Speaker identifications worked equally well. My old M1 Air didn't hang a single bit while the transcription was going on.
---
1. Here's one from a single speaker video (https://www.youtube.com/watch?v=-X6YzlY_8tM): https://pastebin.com/vPVPNnne
2. A shorter with up to 4 different speakers and mixed, complex scene/narration changes (https://www.youtube.com/watch?v=4tASl0auPOg): https://pastebin.com/iHZZD8Qe
--
My zsh shorthand: `alias transcribe="yapsnap --timestamps --diarize "`
Reading this means a lot, thank you! Even faster version in the coming days, stay tuned and PRs welcome!
Awesome.. and thanks for building this. Works superb.
Am I a bit thick, but first we created this amazing way to transfer any text very cheaply and fast over network, then we (well, I think it was Meta and Google) decided that no, everything must be a video, then we added subtitles and AI-transcriptions to those videos and now we just dowload transcriptions of those videos presumably to feel LLM to make summaries of them in order to… Read. Them.
I think I’m gonna go read a book.
Good point! I haven't found a faster way to consume info than reading. But depends on the type of learner you are (visual, auditory, hands-on/interactive, etc)
So, this project consists of a ~175 line README and a ~500 line Python program that glues yt-dlp and Kroko together. Neat.
I guess if it encourages you to install and figure out how to use ffmpeg, yt-dlp, kroko, numpy, and onnx that's a good thing. Sometimes just knowing a thing is possible is a huge benefit.
I see the value as a centralized anti-content-blocker.
This repo is now a good way to centralize hacks around the sure-to-come blockers those platforms will add to prevent download.
Just like uBlockOrigin was a way to centralize all the "just run this greasemonkey script" comments, I can see this getting a huge following for people who really value transcriptions.
I appreciate the perspective! higher ceiling than I'd put on it, but if it gets there awesome. PRs welcome!
thank you. You nailed the actual value, that's right. The real win is just knowing you can do this on a laptop CPU, offline, no GPU or cloud bill. There are tiny done-for-you details, like rescaling token timestamps back to real time after the atempo speedup so --timestamps doesn't lie to you, but they are minor.
Why the choice of Kroko over something like parakeet-tdt-0.6b-v3, which is also faster than realtime on CPU?
Kroko models are more accurate and their size is just a hundred megabytes compared to parakeet (2.5 gigabytes in default fp32)
Do you have a link to results confirming this? Kroko does not seem to be on the Open ASR Leaderboard. Parakeet has an average WER of 6.32 across several common datasets.
Kroko's website says benchmarks aren't formalized yet. FWIW, this url says 5% WER for English [0]. though it doesn't specify the dataset, so not directly comparable to Parakeet's 6.32 on the Open ASR Leaderboard
Best way to judge is to try it on your own audio
[0] https://huggingface.co/hudaiapa88/sherpa-stt-onnx
Works extremely well. Command to install on Debian 13:
sudo apt update && sudo apt install -y ffmpeg python3-pip python3-venv && git clone https://github.com/kouhxp/yapsnap.git && cd yapsnap && python3 -m venv ~/yapsnap-venv && source ~/yapsnap-venv/bin/activate && pip install --upgrade pip && pip install .
On a 32GB ThinkPad X13, a 21 minute YouTube video was processed by yapsnap under 2 minutes.
Very well done!
thank you!
Very cool, I'm also working on a captioning/subtitling project for the lecture recordings for the university I work at.
My biggest challenge is finding a proper language model that is fast enough and accurate enough since I have to caption about 600 hours of video per week and I preferably want to run all of this on a tiny server (2 cores 4 GB memory). This tool could easily do that with the kroko model but I'll have to test if the accuracy is good enough.
Also in my own scripts I'm using ffmpeg to download just the audio of the videos that I want to caption, which saves a lot of bandwith and speeds up the whole process. As far as I can see this tool doesn't do that, that would be a nice functionality to add, plus an option to turn the output into a working .srt file.
thanks! making a note of the feature request
Tis is very simple and very cool! Just installed it on my Hetzner box where I run a remote controlled local agent so now I can basically chat/email a video link to get a summary and/or ask questions. The only issue was YouTube's PO Token requirement (web/mweb clients refuse to serve formats from datacenter IPs without a valid Proof-of-Origin token.) So I had to find a client that still work without PO Token first. Thanks for sharing!
thank you! good use case, what hetzner box specs have you chosen?
CPX22
Had Claude test it out on 3 videos. Worked at 5-8x realtime. The beauty of it is that it works on all videos, not just the one with transcripts. Combine it with YouTube search and LLM takeaways from transcripts, and you have super-efficient content consumption. There are SaaS products that charge 1 cent per video for those with transcripts. There is a viable product in here somewhere, methinks.
thanks for running it Niraj. I see something similar on my machine, which still surprises me every time lol
Wouldn't it still be more efficient to do GPU transcriptions anyways? is this something we could actually put the effectively useless NPUs to use in modern laptops?
yes GPU is significantly faster, but cpu only lets you do it anywhere - wasm in the browser, any server etc.
NPUs - definitely a good use case for at least part of it, there are ports of whisper that use coreML/ANE with less power and 3x speed of CPU only
Possibly, but you may want to use the GPUs for other things, or have under-utilized CPU-only servers lying around.
I thought ONNX models were only for text-to-speech? How does one tell them apart if I find some files online?
Nice. Can it do speaker diarization?
will work on it, that would be neat. I love pyannote but not happening on CPU at reasonable speeds lol
done! just pushed a new version with CPU diarization
How is this so much faster than even GPU-based whisper?
small, ONNX-optimized models designed specifically for low-latency CPU streaming, so it avoids overhead of large transformer arch and GPU memory transfers
Most of these platforms already have transcriptions built in.
Youtube has transcripts on most videos, not all. The others don't expose them. If you mean the "transcript APIs" for TikTok/IG/X, they are all transcribing audio like yapsnap does. If you have a way to pull native ones, let me know, genuinely curious.
YouTube's is transcribing the audio too. The other do expose them as subtitles as the video is playing.
Yes fair point, asr cached and exposed. I meant to draw the line more on fetchable or not.
How can we transcribe other languages besides English?
Just download the model for your preferred language, all hosted on the Kroko-ASR collection here: https://huggingface.co/Banafo/Kroko-ASR/tree/main Right now you have Dutch, French, Portuguese, Spanish, German, Italian, Swedish, Swiss German, Hebrew, and Turkish. Grab the one that matches your audio, point yapsnap at it with --model (or set KROKO_MODEL), and you're set!
Was hoping for CJK languages but I don't see any there. Thanks anyway
Added Diarization / Speaker Separation that is fast and CPU only. Thank you all for the great feedback and support. PRs welcome!
dubaiamir1111
Now make it distinguish speakers and we really have something. As far as I know, that's significantly harder though.
done! new version separates speakers on CPU fast
I'm surprised!
in the roadmap!