: I added the recognized text so that I can confirm what was translated. I also moved my-type-with-hint to learn-lang-type-with-hint.
When I'm writing a journal entry in French, I
sometimes want to translate a phrase that I can't
look up word by word using a dictionary.
Instead of switching to a browser, I can use an
Emacs function to prompt me for text and either
insert or display the translation.
The plz library makes HTTP requests slightly
neater.
I think it would be even nicer if I could use speech synthesis, so I can keep it a little more separate from my typing thoughts. I want to be able to say "Okay, translate …" or "Okay, … in French" to get a translation. I've been using my fork of natrys/whisper.el for speech recognition in English, and I like it a lot. By adding a function to whisper-after-transcription-hook, I can modify the intermediate results before they're inserted into the buffer.
But that's too easy. I want to actually type things myself so that I get more practice. Something like an autocomplete suggestion would be handy as a way of showing me a hint at the cursor. The usual completion-at-point functions are too eager to insert things if there's only one candidate, so we'll just fake it with an overlay. This code works only with my whisper.el fork because it supports using a list of functions for whisper-insert-text-at-point.
(defunmy-whisper-maybe-type-with-hints (text)
"Add this function to `whisper-insert-text-at-point'."
(let* ((hint (and text (org-find-text-property-in-string 'type-hint text)))
(original (and text (org-find-text-property-in-string 'type-original text))))
(if hint
(progn
(learn-lang-type-with-hint hint original)
nil)
text)))
Here's a demonstration of me saying "Okay, this is a test, in French.":
Screencast of using speech recognition to translate into French and provide a hint when typing
First, I need a Python server that can print out events when it notices the start or stop of a speech segment. If I print out the timestamps, I might be able to cross-reference it someday with interestingthings. For now, even just paying attention to the end of a segment is enough for what I want to do.
Python script for printing out events
import sounddevice as sd
import numpy as np
import torch
import sys
from datetime import datetime, timedelta
SILENCE_DURATION= 500
SAMPLING_RATE= 16000
CHUNK_SIZE= 512
model, utils= torch.hub.load(repo_or_dir='snakers4/silero-vad',
model='silero_vad',
force_reload=False)
(get_speech_timestamps, save_audio, read_audio, VADIterator, collect_chunks) = utils
vad_iterator= VADIterator(model, threshold=0.5, min_silence_duration_ms=SILENCE_DURATION)
stream_start_time=Nonedefformat_iso_with_offset(offset_seconds):
if stream_start_time isNone:
return"PENDING"event_time= stream_start_time + timedelta(seconds=offset_seconds)
return event_time.astimezone().isoformat(timespec='milliseconds')
defaudio_callback(indata, frames, time, status):
global stream_start_time
if status:
print(status, file=sys.stderr)
if stream_start_time isNone:
stream_start_time= datetime.now()
tensor_input= torch.from_numpy(indata.copy()).flatten()
speech_dict= vad_iterator(tensor_input, return_seconds=True)
if speech_dict:
if"start"in speech_dict:
print(f"START {format_iso_with_offset(speech_dict['start'])}", flush=True)
if"end"in speech_dict:
print(f"END {format_iso_with_offset(speech_dict['end'])}", flush=True)
try:
with sd.InputStream(samplerate=SAMPLING_RATE,
channels=1,
callback=audio_callback,
blocksize=CHUNK_SIZE):
whileTrue:
passexceptKeyboardInterrupt:
print("\nStopping...")
I've mode the code to speech-input-vad.el, so now I can use speech-input-vad-start to start monitoring speech events.
Because I added Pulse properties to the process environment, I can easily use epwgraph to rewire the input so that it gets the input from my VirtualMicSink instead of the default system audio device. (Someday I'll figure out how to specify that as the input automatically.)
Now I can press my shortcut for my-whisper-continue to start the process. As I keep talking, it will continue to record. When I pause for more than a second between sentences, then it will send that chunk to the server for transcription without me having to press another button, while still listening for more speech.
How is this different from the streaming approach that many real-time speech recognition services offer? I think this gives me a bit more visibility into and control of the process. For my personal use, I don't need to have everything processed as quickly as possible, and I'm not trying to replicate live captions. I just want to be able to look back over the last five minutes to try to remember what I was talking about. I usually have a lot of quiet time as I think through my next steps, and it's fine to have it catch up then. I also like that I can save time-stamped audio files for later processing, divided according to the speech segments. Those might be a little bit easier to work with when I get around to compositing them into a video.
I want to be able to talk out loud and have the ideas go into Emacs. I can do this in a number of different ways:
I briefly demonstrated a step-by-step approach with natrys/whisper.el with a single file. I press a keyboard shortcut to start the recording, another shortcut to stop the recording, and it transcribes it in the background. But the way whisper.el is set up is that if I press the keyboard shortcut to start recording again it will offer to interrupt the transcription process, which is not what I want. I want to just keep talking and have it process results as things come in.
What I've just figured out is how to layer a semi-continuous interface for speech recognition on top of whisper.el so that while it's processing in the background, I can just press a keyboard shortcut (I'm using numpad 9 to call my-whisper-continue) to stop the previous recording, queue it for processing, and start the next recording. If I use this keyboard shortcut to separate my thoughts, then Whisper has a much easier time making sense of the whole sentence or paragraph or whatever, instead of trying to use the sliding 30 second context window that many streaming approaches to speech recognition try to use.
Question: Did you fix the keyboard delay you've got while speech catches what you're saying?
Sometimes, when the speed recognition kicks in, my computer gets busy. When my computer gets really busy, it doesn't process my keystrokes in the right order, which is very annoying because then I have to delete the previous word and retype it. I haven't sorted that out yet, but it seems like I probably have to lower the priority on different processes. On the plus side, as I mentioned, if I dictate things instead of typing them, then I don't run into that problem at all.
Also, other notes on delays: The continuous speech recognition via Google Chrome shows up fairly quickly, but it's not very precise, and it doesn't have punctuation. Even if there's a little bit of a delay, as long as I press the my-whisper-continue shortcut after each thought, then I can get that text into my Emacs buffer using the nicer transcription from my selected model. There is going to be a bit of a delay for that one because it gets processed at the end of the thought. Also, I need to start thinking in complete sentences instead of just adding one cause after the other as my brain goes on all of these tangents. I think it's pretty promising. There's the continuous speech recognition via Google Chrome if I don't mind the lower accuracy and lack of punctuation, and I can still get the pretty version on the other side.
Why talk out loud? I liked the Bookclub Tapas presentation that Maddie Sullivan did at EmacsConf 2025. Talking out loud helps me be a lot more verbose about what I'm saying, compared to typing things out or even like having to switch to my notes or interrupting my screen with an Org capture buffer. Of course I want to clean that up for putting into a blog post, but given that my life still sometimes has random interruptions from a kiddo who must have my attention at that very minute, having that kind of record that I can at least try to reread afterwards to reconstruct what I was thinking about sounds like it might be helpful.
Still, making sense out loud is hard. I'm not actually used to talking to people that much now. This is probably a good reason for me to experiment with streaming more. Then I get the practice in talking out loud, there are backup recordings, and people can ask questions when things are unclear.
Of course, sometimes the text doesn't quite make sense because of the speech recognition errors. I can usually figure it out from the context. I save the audio as well so that I can go back and listen to it again if I really need to.
Anyway, here's the code for sending the current recording to whisper in the background and starting another recording. It assumes a lot about how things are set up. For example, I'm only testing this with a local speaches server instead of whisper.cpp. You might need to look at my other speech related configuration blog posts and sections in order to make sense of it.
Code for queuing whisper.el requests to a local server
(defvarmy-whisper--queue nil)
(defunmy-whisper-continue (&optional arg)
"Send what we've got so far for transcription and then continue recording.Call with \\[universal-argument] to signal that we can stop."
(interactive"P")
(require'whisper)
(if arg
(my-whisper-done)
(setq whisper--marker (point-marker) whisper--point-buffer (current-buffer))
(when (process-live-p whisper--recording-process)
;; queue only if the last one is not asking for the same file
(unless
(string=
(plist-get
(car
(last my-whisper--queue))
:file)
whisper--temp-file)
(add-to-list
'my-whisper--queue
(list :file whisper--temp-file
:buffer
(format "*result: %s*" (file-name-base whisper--temp-file)))
t))
;; Remove the sentinel; handle results ourselves
(set-process-sentinel whisper--recording-process
(lambda (process event)
(my-whisper-process-queue)))
(interrupt-process whisper--recording-process))
(run-hooks 'whisper-before-transcription-hook)
(whisper--setup-mode-line :show'recording)
(whisper--record-audio)))
(defunmy-whisper-discard ()
"Ignore the previous recording."
(interactive)
(when (process-live-p whisper--recording-process)
;; Remove the sentinel; handle results ourselves
(set-process-sentinel whisper--recording-process
(lambda (process event)
(when (file-exists-p whisper--temp-file)
(delete-file whisper--temp-file))
(my-whisper-process-queue)))
(interrupt-process whisper--recording-process)))
(defunmy-whisper-discard-and-continue ()
"Ignore the previous recording and continue."
(interactive)
(if (process-live-p whisper--recording-process)
(progn;; Remove the sentinel; handle results ourselves
(set-process-sentinel whisper--recording-process
(lambda (process event)
(my-whisper-process-queue)
(my-whisper-continue)))
(interrupt-process whisper--recording-process))
(my-whisper-continue)))
(defunmy-whisper-done ()
(interactive)
(when (process-live-p whisper--recording-process)
(add-to-list
'my-whisper--queue
(list :file whisper--temp-file
:buffer
(format "*result: %s*" (file-name-base whisper--temp-file)))
t)
;; Remove the sentinel; handle results ourselves
(set-process-sentinel whisper--recording-process
(lambda (process event)
(my-whisper-process-queue)))
(whisper--setup-mode-line :hide'recording)
(interrupt-process whisper--recording-process)))
(defunmy-whisper-process-queue-result ()
"Process the first part of the queue that already has results."
(while (plist-get (car my-whisper--queue) :results)
(let ((o (pop my-whisper--queue)))
(unless my-whisper-target-markers
(setq whisper--marker (point-marker)
whisper--point-buffer (current-buffer)))
(with-current-buffer (plist-get o :buffer)
(erase-buffer)
(insert (plist-get o :results)))
;; Only works with my fork: https://github.com/sachac/whisper.el/tree/whisper-insert-text-at-point-function
(whisper--handle-transcription-output nil (plist-get o :buffer)))))
(defunmy-whisper-process-queue ()
(let (o)
(while (setq o (seq-find (lambda (o) (and (plist-get o :file)
(not (plist-get o :process))
(not (plist-get o :results))))
my-whisper--queue))
(let* ((headers (list "Content-Type: multipart/form-data"))
(params (list (concat "file=@"
(plist-get o :file))
"temperature=0.0""temperature_inc=0.2""response_format=json"
(concat "model=" whisper-model)
(concat "language=" whisper-language)))
(url (format my-whisper-url-format whisper-server-host whisper-server-port))
(command `("curl""-s"
,url
,@(mapcan (lambda (h) (list "-H" h)) headers)
,@(mapcan (lambda (p) (list "-F" p)) params))))
(with-current-buffer (get-buffer-create (plist-get o :buffer))
(erase-buffer))
(plist-put
o :process
(make-process
:name"whisper-curl":command command
:buffer (plist-get o :buffer)
:coding'utf-8:sentinel
(lambda (process event)
(with-current-buffer (process-buffer process)
(let ((current my-whisper--queue-item))
(when (and (get-buffer (plist-get current :buffer))
(string-equal "finished\n" event))
(with-current-buffer (plist-get current :buffer)
(goto-char (point-min))
(plist-put current :results
(or
(condition-case nil
(gethash "text" (json-parse-buffer))
(error""))
"(error)"))))))
(my-whisper-process-queue-result))))
(plist-put o :command (string-join command " "))
(with-current-buffer (process-buffer (plist-get o :process))
(setq-local my-whisper--queue-item o))))))
(defvar-localmy-whisper--queue-item nil)
(defunmy-whisper-reprocess-queue ()
(interactive)
(setq whisper--marker (point-marker) whisper--point-buffer (current-buffer))
(mapc (lambda (o)
(when (process-live-p (plist-get o :process))
(kill-process (plist-get o :process)))
(when (get-buffer (plist-get o :buffer))
(kill-buffer (plist-get o :buffer)))
(plist-put o :process nil)
(plist-put o :results nil))
my-whisper--queue)
(my-whisper-process-queue))
(defunmy-whisper-clear-queue ()
(interactive)
(mapc (lambda (o)
(when (process-live-p (plist-get o :process))
(kill-process (plist-get o :process)))
(when (get-buffer (plist-get o :buffer))
(kill-buffer (plist-get o :buffer)))
(plist-put o :process nil)
(plist-put o :results nil))
my-whisper--queue)
(setq my-whisper--queue nil))
(keymap-global-set "<kp-9>"#'my-whisper-continue)
(keymap-global-set "<kp-8>"#'my-whisper-discard-and-continue)
(keymap-global-set "C-<kp-9>"#'my-whisper-done)
I was curious about parakeet because I heard that it was faster than Whisper on the HuggingFace leaderboard. When I installed it and got it running on my laptop (CPU only, no GPU), it seemed like my results were a little faster than whisper.cpp with the large model, but much slower than whisper.cpp with the base model. The base model is decent for quick dictation, so I got curious about other backends and other models.
In order to try natrys/whisper.el with other backends, I needed to work around how whisper.el validates the model names and sends requests to the servers. Here's the quick and dirty code for doing so, in case you want to try it out for yourself.
(defvarsacha-whisper-url-format"http://%s:%d/transcribe")
;;;###autoload
(defunsacha-whisper--transcribe-via-local-server ()
"Transcribe audio using the local whisper server."
(message "[-] Transcribing via local server")
(whisper--setup-mode-line :show'transcribing)
(whisper--ensure-server)
(setq whisper--transcribing-process
(whisper--process-curl-request
(format sacha-whisper-url-format whisper-server-host whisper-server-port)
(list "Content-Type: multipart/form-data")
(list (concat "file=@" whisper--temp-file)
"temperature=0.0""temperature_inc=0.2""response_format=json"
(concat "model=" whisper-model)
(concat "language=" whisper-language)))))
;;;###autoload
(defunsacha-whisper--check-model-consistency () t)
Update: After a lot of rebooting, I figured out how to get the Nvidia P1000 GPU on my Lenovo P52 to work for up to the Whisper medium models. Large models don't fit in the 4GB of VRAM it has. Here are the benchmarking results:
That probably means I can upgrade to using the small model on a regular basis. For large models, I can use the command-line tool, since the overhead of loading the model is probably small compared to the processing time. I could also run two Speaches servers (one on CPU and one on GPU), but I don't think I'll need that level yet.
I like how speaches lets me switch models on the fly, so maybe I can use small.en generally and switch to a different model when I want to try dictating in French. Here's how I've set it up to use the server.
Benchmark notes: I have a Lenovo P52 laptop (released 2018) with an Intel Core i7-8850H (6 cores, 12 threads; 2.6 GHz base / 4.3 GHz turbo) with 64GB RAM and an SSD. It has a soldered-on Nvidia P1000 with 4GB of VRAM.
: I now use Orukeet as the speech recognition model because it's faster.
: Major change: I switched to my fork of natrys/whisper.el so that I can specify functions that change the window configuration etc.
: Change main function to sacha-whisper-run, use seq-reduce to go through the functions.
: Added code for automatically capturing screenshots, saving text, working with a list of functions.
: Added demo, fixed some bugs.
: Added note about difference from MELPA package, fixed :vc
I want to get my thoughts into the computer quickly, and talking might be a good way to do some of that. whisper.el gives me a convenient way to interact with a server compatible with OpenAI Whisper's API. (Experimenting with Orukeet for speech recognition) Note: This is not the same whisper package as the one on MELPA. Here is how I have it set up for reasonable performance on my Lenovo P52 with just the CPU, no GPU.
I've bound <f9> to the command whisper-run in Emacs, but I use my window manager's custom shortcuts to map <kp-0> to execute /home/sacha/bin/xdotool-emacs key --clearmodifiers F9, which sends <f9> to Emacs. (file:///home/sacha/bin/xdotool-emacs) That makes it a shortcut that works even in other applications.
I press <kp-0> to start recording, talk, and then press <kp-0> to stop recording. By default, it inserts the text into the buffer at the current point. I've set whisper-return-cursor-to-start to nil so that I can keep going.
Update : Il peut aussi comprendre le français. I used to have code to set the language in case the auto-recognition doesn't get it right, but now I'm using Orukeet as the speech recognition engine now, so it doesn't take a language setting.
;; Only works with my tweaks to whisper.el;; https://github.com/sachac/whisper.el/tree/whisper-insert-text-at-point-function
(with-eval-after-load'whisper
(setq whisper-insert-text-at-point
'(sacha-whisper-scratch-that
sacha-whisper-handle-commands
sacha-whisper-save-text
sacha-learn-lang-notice-bingo-words
; sacha-whisper-save-to-file
sacha-save-to-kill-ring-after-current
sacha-whisper-maybe-expand-snippet
sacha-speech-input-quantified-track
sacha-whisper-maybe-type
sacha-whisper-maybe-type-with-hints
sacha-whisper-insert
sacha-whisper-reset)))
(defvarsacha-whisper-last-annotation nil "Last annotation so we can skip duplicates.")
(defvarsacha-whisper-skip-annotation nil)
(defvarsacha-whisper-target-markers nil "List of markers to send text to.")
;;;###autoload
(defunsacha-whisper-insert (text)
(let ((markers
(cond
((null sacha-whisper-target-markers)
(list whisper--marker)) ; current point where whisper was started
((listp sacha-whisper-target-markers)
sacha-whisper-target-markers)
((markerp sacha-whisper-target-markers)
(list sacha-whisper-target-markers))))
(orig-point (point))
(orig-buffer (current-buffer)))
(when text
(mapcar (lambda (marker)
(with-current-buffer (marker-buffer marker)
(save-restriction
(widen)
(when (markerp marker) (goto-char marker))
(when (and (derived-mode-p 'org-mode) (org-at-drawer-p))
(insert "\n"))
(whisper--insert-text
(concat
(if (looking-back "[ \t\n]\\|^")
""" ")
(string-trim text)))
;; Move the marker forward here
(move-marker marker (point)))))
markers)
(when sacha-whisper-target-markers
(goto-char orig-point))
nil)))
;;;###autoload
(defunsacha-whisper-maybe-type (text)
"If Emacs is not the focused app, simulate typing TEXT.Add this function to `whisper-insert-text-at-point'."
(when text
(if (frame-focus-state)
text
(make-process :name"xdotool":command
(list "xdotool""type"
text))
nil)))
Here's an idea for a function that saves the recognized text with a timestamp.
(defvarsacha-whisper-notes"~/sync/stream/narration.org")
;;;###autoload
(defunsacha-whisper-save-to-file (text)
(when text
(let ((link (org-store-link nil)))
(with-current-buffer (find-file-noselect sacha-whisper-notes)
(goto-char (point-max))
(insert "\n\n" (format-time-string "%H:%M ") text "\n" (if link (concat link "\n") ""))
(save-buffer)
(run-at-time 0.5 nil (lambda (text) (message "Saved to file: %s" text)) text)))
text))
Let's add it to the kill ring, but not at the most recent slot. That way, it's easy for me to get to, but it doesn't overwrite anything that I've copied.
;;;###autoload
(defunsacha-save-to-kill-ring-after-current (text)
"Save TEXT to the kill ring, but not at the top spot."
(when text
(if kill-ring
(let ((temp (pop kill-ring)))
(kill-new text)
(push temp kill-ring))
(kill-new text)))
text)
I think I've just figured out my Pipewire setup so
that I can record audio in OBS while also being
able to do speech to text, without the audio
stuttering. qpwgraph was super helpful
for visualizing the Pipewire connections and fixing them.
Screencast of using whisper.el to do speech-to-text into the current buffer, clocked-in task, or other function
Transcript
0:00Inserting into the current buffer
Here's a quick demonstrationof using whisper.el to log notes.
0:13Inserting text and moving on
I can insert text into the current bufferone after the other.
0:31Clocking in
If I clock into a task,I can add to the end of that clocked in taskusing my custom codeby pressing C-<f9>or whatever my shortcut was.I can do that multiple times.
1:05Logging a note from a different file
I can do that while looking at a different file.
1:15I can look at an info page
I can do it looking at an info page, for example,and annotations will include a linkback to whatever I was looking at.
1:33Adding without an annotation (C-u)
I just added an optional argumentso that I can also capture a notewithout saving an annotation.That way, if I'm going to say a lot of thingsabout the same buffer,I don't have to have a lot of linksthat I need to edit out.
2:42Saving to a different function
I can also have it save to a different function.
And then I define a global shortcut in KDE that runs xdotool-emacs:
So now I can dictate into other applications or save into Emacs.
Which suggests of course that I should get it working with C-f9 as well, if I can avoid the keyboard shortcut loop…
