Spaces:
Runtime error
Runtime error
Commit History
-
Update app.py
47bd609 -
customization
c11b3a5
-
Fix pad_timestamps
02253c6 -
Return results, not last seconds
bcee468 -
Process VAD in chunks of up to 1 hour
b604ea1
-
Remove unused imports
2ea4ed3
-
Fix unit test
25df8a0 -
Change the default merge window to 5.
This is effectively 7 seconds, as each speech section is also padded with one second.b1d4eff -
Fix merging of VAD segments
Also expand VAD segments rather than running whisper on non-speech segments. That way, whisper will use the previous section as a prompt, in case there is speech in the non-speech section after all.8781620 -
Move implementation to src folder
c52f09b -
Handle OS error 36 - File name too long
On many Linux systems, the maximum file name lenght is 255 bytes. If the title contains UTF-8 characters, each character may require more than 1 byte in the file name, causing a title of 100 characters to actually exceed the 255 byte limit. We handle this error and then use a much stricter file name format.11c4fb6
-
Do not run whisper on very short VAD segments
89bd811 -
Fix options.md
a95d6a8 -
Add max merge size in VAD
Also make this and the VAD merge window configurable.084aa80 -
Import tensorflow if it exists
This is necessary to avoid the following error when TensorFlow is present: "SystemError: google/protobuf/pyext/descriptor.cc: 358: bad argument to internal function" More details here: https://github.com/tensorflow/tensorflow/issues/48797ea7f8cc
-
Transcribe non-speech areas too in "silero vad"
The old version is now in "silero-vad-skip-gaps". This may introduce more noise in the transcript, but some of it will be correct as well.bc0cb58 -
Add tourchaudio as a requirement
55fc070 -
Remove type definition for callback to support old Python
ff19da3 -
Add a simple "VAD" alternative - periodic VAD
This is a very simple alternative to using Silero Vad, where every 5 minutes (0:00 - 5:00, 5:00 - 10:00, etc.) is marked as a "speech" segment, and whisper is then run on each segment individually. The upside is that all potential dialogue will be sent to Whisper for transcribing, but it is also more likely to get into an infinite sentence loop (although limited to each 5 minute segment). Creating an artifical break every 5 minutes may also break up a sentence accidentally, causing the sentence (or a word) to be incorrectly transcribed.bbbf06e -
Disable buffering in docker container
This allows us to view the logs in real time0d8d833 -
Add support for selecting a VAD
The VAD (Voice Activity Detector) is used to detect time segments where there is speech (more than 250ms), and only run whisper on these segments. This prevents Whisper getting stuck in a loop, producing the same sentence over and over (which usually happens after a long continuous sequence of no speech). A secondary benefit is that the time synchronization issues that sometimes are present, will not carry over each detected time segment. One slight issue, however, may be if these time sequences are very short, and prevent Whisper from using previous text as prompt for context. To mitigate this somewhat, the detected time segments with speech are padded by 1 second before, and 4 seconds after, and then merged if they overlap. Finally, the VAO model's threshold is set to 30% instead of 50%.f288ceb
-
Add cache information
aa22372 -
Update README
a34af3d -
Add support for docker
a14ac5a -
Fix detection of JP/ZH when calculating max line length
ba1ed2f
-
Fix YouTube video length restriction
8f5637c
-
Improve description
38cc8a7 -
Limit video length before downloading from YouTube
fdd892b -
Let max line width depend on the language
6a308c6 -
Wrap lines in SRT and VTT files that exceed 47 characters
1a68fc3
-
Create a new app-network version
08d2e96 -
Expose app-local to the local network
d5154e9 -
Fix SRT generation
15219b4 -
Also delete uploaded files to save space
533d92e -
Increase audio file limit to 10 minutes
4514e2e -
Add support for downloading files from YouTube using yt-dlp
8d120bf -
Add support for downloading the results.
3fadc6e -
Increase limit to 5m
123e71d
-
Rename app-full.py to app-local.py
Kristian Stangeland4b1adcf -
Update requirements and README
Kristian Stangelandbf299de -
Add support for sharing
de7a3ac -
Improve README
02edf9e -
Make it easier to run with no audio file restrictions
71950a8 -
Cleanup code
93c4867
-
Actually limit to 120s
68a16d3 -
Limit audio files to 120s
7ce6041 -
Initial commit
05a2178 -
initial commit
23c3153