Using WhisperX to get word-level timestamps for audio editing with Emacs and subed-record
Posted: - Modified: | audio, emacs, subed- : Split whisperx to pass one file per call.
- : Removed
highlight_words, mademax_line_widthuse the environment variable if specified. - : Actually, WhisperX makes a JSON with word-level timing data, so let's use that instead.
I'm gradually shifting more things to this Lenovo P52 to take advantage of its newer processor, 64 GB of RAM, and 2 TB drive. (Whee!) One of the things I'm curious about is how I can make better use of multimedia. I couldn't get whisper.cpp to work on my Lenovo X230T, so I mostly relied on the automatic transcripts from Google Recorder (with timestamps generated by aeneas) or cloud-based transcription services like Deepgram.
I have a lot of silences in my voice notes when I think out loud. whisper.cpp got stuck in loops during silent parts, but WhisperX handles them perfectly. WhisperX is also fast enough for me to handle audio files locally instead of relying on Deepgram. With the default model, I can process the files faster than real-time:
| File length | Transcription time |
|---|---|
| 42s | 17s |
| 7m48s | 1m41s |
I used this command to get word-level timing data. (Experimenting with options from this post)
MAX_LINE_WIDTH="${MAX_LINE_WIDTH:-50}" MODEL="${MODEL:-large-v2}" for FILE in "$@"; do ~/vendor/whisperx/.venv/bin/whisperx --model "$MODEL" --align_model WAV2VEC2_ASR_LARGE_LV60K_960H --compute_type int8 --print_progress True --max_line_width $MAX_LINE_WIDTH --segment_resolution chunk --max_line_count 1 --language en "$FILE" done
Among other things, it makes a text file that looks like this:
I often need to... I sometimes need to replace or navigate by symbols. Casual symbol overlays a new package that adds those shortcuts so that I don't have to remember the other keywords for them.
and a JSON file that looks like this:
{"segments": [{"start": 0.427, "end": 7.751, "text": " I often need to... I sometimes need to replace or navigate by symbols.", "words": [{"word": "I", "start": 0.427, "end": 0.507, "score": 0.994}, {"word": "often", "start": 0.587, "end": 0.887, "score": 0.856}, {"word": "need", "start": 0.987, "end": 1.227, "score": 0.851}, {"word": "to...", "start": 1.267, "end": 1.508, "score": 0.738}, {"word": "I", "start": 4.329, "end": 4.429, "score": 0.778}, ...]}, ...]}
Sometimes I just want the text so that I can use an audio braindump as the starting point for a blog post or for notes. WhisperX is way more accurate than Google Recorder, so that will probably be easier once I update my workflow for that.
Sometimes I want to make an edited audio file that sounds smooth so that I can use it in a podcast, a video, or some audio notes. For that, I'd like word-level timing data so that I can cut out words or sections. Aeneas didn't give me word-level timestamps, but WhisperX does, so I can get the time information before I start editing. I can extract the word timestamps from the JSON like this:
(defun my-subed-word-tsv-from-whisperx-json (file) (interactive "FJSON: ") (let* ((json-array-type 'list) (json-object-type 'alist) (data (json-read-file file)) (filename (concat (file-name-sans-extension file) ".tsv")) (base (seq-mapcat (lambda (segment) (seq-map (lambda (word) (let-alist word (list nil (and .start (* 1000 .start)) (and .end (* 1000 .end)) .word))) (alist-get 'words segment))) (alist-get 'segments data))) (current base) (last-end 0)) ;; numbers at the end of a sentence sometimes don't end up with times ;; so we need to fix them (while current (unless (elt (car current) 1) ; start (setf (elt (car current) 1) (1+ last-end))) (unless (elt (car current) 2) (setf (elt (car current) 2) (1- (elt (cadr current) 1)))) (setq last-end (elt (car current) 2) current (cdr current))) (subed-create-file filename base t 'subed-tsv-mode) (find-file filename)))
Here's my old code for parsing the highlighted VTT or SRT files that underline each word:
(defun my-subed-load-word-data-from-whisperx-highlights (file) "Return a list of word cues from FILE. FILE should be a VTT or SRT file produced by whisperx with the --highlight_words True option." (seq-keep (lambda (sub) (when (string-match "<u>\\(.+?\\)</u>" (elt sub 3)) (setf (elt sub 3) (match-string 1 (elt sub 3))) sub)) (subed-parse-file file))) (defun my-subed-word-tsv-from-whisperx-highlights (file) (interactive "FVTT: ") (with-current-buffer (find-file-noselect (concat (file-name-nondirectory file) ".tsv")) (erase-buffer) (subed-tsv-mode) (subed-auto-insert) (mapc (lambda (sub) (apply #'subed-append-subtitle nil (cdr sub))) (my-subed-load-word-data-from-whisperx-highlights file)) (switch-to-buffer (current-buffer))))
I like to use the TSV format for this one because it's easy to scan down the right side. Incidentally, this format is compatible with Audacity labels, so I could import that there if I wanted. I like Emacs much more, though. I'm used to having all my keyboard shortcuts at hand.
0.427000 0.507000 I 0.587000 0.887000 often 0.987000 1.227000 need 1.267000 1.508000 to... 4.329000 4.429000 I 4.469000 4.869000 sometimes 4.950000 5.170000 need 5.210000 5.410000 to 5.530000 6.090000 replace
Once I've deleted the words I don't want to include, I can merge subtitles for phrases so that I can keep the pauses between words. A quick heuristic is to merge subtitles if they don't have much of a pause between them.
(defvar my-subed-merge-close-subtitles-threshold 500) (defun my-subed-merge-close-subtitles (threshold) "Merge subtitles with the following one if there is less than THRESHOLD msecs gap between them." (interactive (list (read-number "Threshold in msecs: " my-subed-merge-close-subtitles-threshold))) (goto-char (point-min)) (while (not (eobp)) (let ((end (subed-subtitle-msecs-stop)) (next-start (save-excursion (and (subed-forward-subtitle-time-start) (subed-subtitle-msecs-stop))))) (if (and end next-start (< (- next-start end) threshold)) (subed-merge-with-next) (or (subed-forward-subtitle-end) (goto-char (point-max)))))))
Then I can use subed-waveform-show-all to tweak the start and end timestamps.
Here I switch to another file I've been editing…
After that, I can use subed-record to compile the
audio into an .opus file that sounds reasonably smooth.
I sometimes need to replace or navigate by symbols. casual-symbol-overlay is a package that adds a transient menu so that I don't have to remember the keyboard shortcuts for them. I've added it to my embark-symbol-keymap so I can call it with embark-act. That way it's just a C-. z away.
I want to make lots of quick audio notes that I can shuffle and listen to in order to remember things I'm learning about Emacs (might even come up with some kind of spaced repetition system), and I'd like to make more videos someday too. I think WhisperX, subed, and Org Mode will be fun parts of my workflow.