Is the point that you only need one tool -- ffmpeg -- to both generate transcripts as well as embed those into a video as opposed to having multiple tools?
This is a 3 part series, the first one discusses the new native whisper integration. And correct, for the first post - the point is to show that you can only use ffmpeg to transcribe and embed subtitles in a video
While there's appeal in having one tool do several things I'm more a fan of the traditional UNIX philosophy that a tool should do one thing, do it extremely well, and allow for chaining of several tools together to achieve a multi-step process.
Whisper will hallucinate on audio segments that don't have any speech. VAD mitigates that. Expect worse results without it, especially on non-English audio.
Do you know if it’s supported on Mac too, with whatever platform specific optimizations like running it on the gpu / with MPS ?
You mean Vulkan? In the blog post there is reference to all vulkan supported platforms
If you mean ffmpeg build with whisper - from memory I didn't see ffmpeg-builds for mac, so you will probably need to compile yourself
Is the point that you only need one tool -- ffmpeg -- to both generate transcripts as well as embed those into a video as opposed to having multiple tools?
This is a 3 part series, the first one discusses the new native whisper integration. And correct, for the first post - the point is to show that you can only use ffmpeg to transcribe and embed subtitles in a video
While there's appeal in having one tool do several things I'm more a fan of the traditional UNIX philosophy that a tool should do one thing, do it extremely well, and allow for chaining of several tools together to achieve a multi-step process.
I tend to agree. The thing I like most about version 8 is actually pad_cuda - nice performance boost for resizing video with an Nvidia GPU
Author here, available for questions
"Lenovo laptop with Nvidia RTX 4040" 4060?
Correct. I fixed the typo
This is great I’ll have to give it a shot
Whisper will hallucinate on audio segments that don't have any speech. VAD mitigates that. Expect worse results without it, especially on non-English audio.