| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
At the core of the STT, there's a loop which uses Whisper to convert
audio into a transcript. As you say something, whisper sees growing
fragments of your sentence:
t0: "Hell"
t1: "Hello"
t2: "Hello, world!"
So we need some algorithm which takes these fragments and
accumulates them into an ever-growing transcript.
Previously I did this with fuzzy string matching. I'd find the region
where the two transcripts overlap and edit the two together to produce a
longer transcript. The big problem is that if there's no overlap, it's
not clear whether whisper radically changed its mind as to what was
said, or whether the user paused for a long time before saying
something new. So I'd have to reset the growing transcript.
Now I get the timestamps from Whisper and wait for it to give me the
same 3 transcripts for the last utterance. Once the transcript
stabilizes like this, I commit the text. This enables a temporally
stable, ever-growing transcript that's also quite accurate.
To prevent a latency regression, I also introduce the notion of "preview
text", which is a preview of an utterance that has not yet stabilized.
These previews do not contribute to the ever-growing transcript, but do
get fed through the rest of the app, so they show up in-game / in OBS.
Once they eventually stabilize, they get committed to the ever-growing
transcript.
This change is lightly tested!
|
| |
|
|
|
|
|
|
| |
Use Const-me/Whisper to perform transcription. This implementation is
vastly more efficient: CPU usage, memory usage, and VRAM usage are all
dramatically reduced. It's slightly less accurate when comparing the
same model (due to the lack of beam search decoding), but since you can
use larger models, the impact is largely a wash.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, paths containing spaces would be interpreted by python's argument
parser as multiple separate arguments, causing it to fail. Now we escape paths
inside PythonWrapper using std::quoted().
* Improve PII filtering. Python output would contain multiple path separators
(like C:\\Users\\foo\\), defeating the PII regex.
* Silence compiler warning in PII filter.
* Document usability improvements.
* Transcription layer exponential backoff goes to ~infinity when paused.
This is a hack, since we really don't need to transcribe at all when paused,
but it lets us keep the code simple. Good enough until the next rewrite.
* Shader only samples background when necessary.
* Limit matchStrings() print()s to DEBUG mode
|
| |
|
|
|
|
|
|
| |
Bump up recording window to 28 seconds. This helps a lot with long-form
transcription tasks, s.a. transcribing an audiobook.
We should expose this as a parameter, since at 10s the transcription delay is
typically 300ms, while at 28s it's typically 1.1-1.2s.
|
|
|
GUI can now download all TaSTT dependencies and install them into a
virtual environment.
* Add buttons to check embedded python version & install dependencies
* Add class to wrap interacting with embedded Python
* Put all TaSTT python scripts into a folder
|
