Levi Melamed

> would it be possible in theory to enable word_level timestamps through faster_whisper and patch them into the segments? Yes. The timestamp tokens are being [filtered out during decoding](https://github.com/m-bain/whisperX/blob/78dcfaab51005aa703ee21375f81ed31bc248560/whisperx/asr.py#L69). You...

> Maybe `--chunk_size` is what you need. > > https://github.com/m-bain/whisperX/blob/f8cc46c6f7fa3b8509bc6aa04cdf4a62a702bb42/whisperx/transcribe.py#L44 Wouldn't this cause worse transcription results, since the context is lost with smaller chunks of audio? Seems like it would...

I'm getting that error when I attempt to call `count_tokens()` with a large string.

Instead of copying the file to the worker, you can transfer a reference to it like so: `await this.ffmpeg.mount("WORKERFS", { files: [file]}, '/some-dir');` [See here](https://emscripten.org/docs/api_reference/Filesystem-API.html#filesystem-api-workerfs)

Some interesting behavior I found while testing scanForColors: ``` tinycolor.scanForColors("background-position: 100% 50%;"); // finds a match: ["100"] ``` The culprit is the hex3 matcher, which looks like this once you...

A few more strings that are returning matches: ``` z-index: 100; margin: 0 !important; // finds the word "tan" ```

I think requiring a preceding `#` makes sense. I see it as being no different than the `rgba?` which precedes rgb colors. Without it, you are prone to all sorts...

How about an optional object parameter `scanForColors(text, { css: true })` which defaults to false? Based on the parameter, CSS syntax enforcements could be toggled on/off.

Any updates on this? Experiencing the same issue on `8.2.32` on `M3 Max`. CPU validation matches the results of Cuda validation, MPS gives much worse results.

It looks like `whispers2t` does not use the previous segment transcription as context. This is the same with `WhisperX`. Would be interesting to see WER benchmarks alongside the performance, especially...