This tool allows you to use AI models to generate subtitles from only audio, then match the subtitles to an accurate text, like a book. It requires a modern GPU with decent VRAM, CPU, and RAM.

Current State: The transcript will be extremely accurate. The timings will be mostly accurate, but may come late or leave early. The currently used library for generating those offsets is the best I've found so far that works stably, but leaves much to be desired. See the video at the bottom for such an example.

I'm looking forward to being able to run more accurate models to fix this in the future.

Currently supports unix based OS's like Ubuntu 20.04 on WSL2.

1. Install ffmpeg and make it available on the path

2. Use python 3.9.9

3. pip install stable-ts

4. Be able to run the docker image for m4b-tool. Trust me, you want the improved codec's that are included in the docker image. I tested both and noticed a huge drop in sound quality without them.

Primarily I'm using this for syncing audiobooks to their book script. So while you could use this for video files, I'm not doing that just yet. Note: I'm using the term "splitted" because that's what m4b refers to as the split files.

1. git clone https://github.com/kanjieater/AudiobookTextSync.git
2. Make sure you run any commands that start with ./ from the project root, eg after you clone you can run cd ./AudiobookTextSync
3. Setup the folder. Create a folder to hold a single media file (like an audiobook). Name it whatever you name your media file, eg Arslan Senki 7, this is what should go anywhere you see me write <name>
4. Put the script in place: ./<name>/script.txt. Everything in this file will show up in your subtitles. So it's important you trim out excess (table of contents, character bios that aren't in the audiobook etc)
5. You need both the audiobook as a full m4b (technically other formats would work), AND the split parts. As long as you have one, you can easily get the other. You could technically only use the full single file, but you will most-likely run out of ram for longer works. See Whisper ~13GB Memory Usage Issue for 19hr Audiobook. By using small splits, we can have more confidence the Speech To Text analysis won't get killed by an Out Of Memory error.
6. Split files should be ./<name>/<name>_splitted/.If you have the full audiobook as a m4b, you can split it into chapters using ./split.sh "<full folder path>". eg ./split.sh "/mnt/d/Editing/Audiobooks/かがみの孤城/"
7. Single media file should be in ./<name>/<name>.m4b. If you have the split audiobook as m4b,mp3, or mp4's you can run ./merge.sh "<full folder path>", eg ./merge.sh "/mnt/d/Editing/Audiobooks/ｍｅｄｉｕｍ霊媒探偵城塚翡翠"
8. If you have the script.txt and ./<name>/<name>_splitted/, you can now run the GPU intense, time intense, and occasionally CPU intense script part. ./run.sh "<full folder path>" eg ./run.sh "/mnt/d/Editing/Audiobooks/かがみの孤城/". This runs each split file individually to get a word level transcript. It then creates a sub format that can be matched to the script.txt. Each word level subtitle is merged into a phrase level, and your result should be a <name>.srt file that can be watched with mpv, showing audio in time with the full book as a subtitle. From there use a texthooker and enjoy.

./split.sh "/mnt/d/Editing/Audiobooks/かがみの孤城/"

./run.sh "/mnt/d/Editing/Audiobooks/かがみの孤城/"

You can also run for a single file. Beware if it's over 1GB/19hr you need as much as 23GB of RAM available. You need two copies of your file. One in "" and one in <full folder path>/splitted_<name>, as described in the How to Use section. The single file will only run if you don't have <name>_splitted folder, otherwise we'll assume you want to use the data from there in parts.

./run.sh "<full folder path>" eg ./run.sh "$(wslpath -a "D:\Editing\Audiobooks\かがみの孤城\\")"

./merge.sh "/mnt/d/Editing/Audiobooks/ｍｅｄｉｕｍ霊媒探偵城塚翡翠"

At this point I would recommend reading from the texthooker instead of a sub. (CTRL+SHIFT+RIGHT in mpv to set offset as the next sub). Then you can see the next line coming in the texthooker, and not be distracted by subtitle jumps.

• Generates subs2srs style deck
• Imports the deck into Anki automatically

The Anki support currently takes your m4b file in <full_folder_path> named <name>.m4b, where <name> is the name of the media, and it outputs srs audio and a TSV file that can is sent via AnkiConnect to Anki. This is useful for searching across GoldenDict to find sentences that use a word, or to merge automatically with custom scripts (more releases to support this coming hopefully).

1. Install ankiconnect add-on to Anki.
2. I recommend using ANKICONNECT as an environment variable. Set export ANKICONNECT=localhost:8755 or export ANKICONNECT="$(hostname).local:8765" in your ~/.zshrc or bashrc & activate it.
3. Make sure you are in the project directory cd ./AudiobookTextSync
4. Install pip install ./requirements.txt (only needs to be done once)
5. Set ANKI_MEDIA_DIR to your anki profile's media path: /mnt/f/Anki2/KanjiEater/collection.media/
6. Run the command below

Command: ./anki.sh "<full_folder_path>"

Example: ./anki.sh "/mnt/d/sync/kokoro/"

If you're using WSL2 there a few networking quirks.

1. Enable WSL2 to talk to your Windows machine. microsoft/WSL#4585 (comment)
2. Set your $ANKICONNECT url to your windows machine url, export ANKICONNECT="http://$(hostname).local:8765". microsoft/WSL#5211
3. Make sure inside of Anki's addon config "webBindAddress": "0.0.0.0", "webBindPort": "8765". 0.0.0.0 binds to all network interfaces, so WSL2 can connect.

curl --header "Content-Type: application/json" \
  --request POST \
  --data '{ "action": "guiBrowse", "version": 6, "params": { "query": "flag:3 is:new -is:suspended -tag:重複 tag:重複3" } }' \
  http://172.18.224.1:8765

You might see various issues while trying this out in the early state. Here are some of the pieces at work in sequence:

1. (not pushed yet) Filter down audio to improve future results - slow & probably not heavy cpu or gpu usage. Heavier on cpu
2. split_run & stable-ts: Starts off heavy on CPU & RAM to identify the audio spectrum
3. stable-ts: GPU heavy & requires lots of vRAM depending on the model. This is the part with the long taskbar, where it tries to transcribe a text from the audio. Currently the default is tiny. Ironically tiny, does a better job of keeping the phrases short, at the cost of accuracy of transcription, which since we are matching a script, doesn't matter. Also it runs 32x faster than large.
4. Merge vtt's for split subs
5. Split the script
6. match the script to the generated transcription to get good timestamps

Besides the other ones already mentioned & installed this project uses other open source projects subs2cia, & anki-csv-importer

https://github.com/gsingh93/anki-csv-importer

https://github.com/kanjieater/subs2cia